That’s the swap I made about two months ago, and the time math is the smaller part of the story. The bigger part is what the twelve-minute version actually returns.

Does this sound familiar when you’re researching content topics?

• You ask Claude (or Perplexity, Grok, ChatGPT - take your pick) what’s trending in your niche.

• You get back a tidy list of buzzwords:”Agentic AI.” “MCP servers.” “The death of prompt engineering.”

• You get a one-line summary on each. Maybe three.

What you don’t get is anything that helps you decide what to actually write about or any meaningful connections like whether any of those buzzwords fit your audience this week, a framework you already own, or if they can become an essay Monday morning without you having to force the angle.

A generic “what’s trending” prompt produces scroll fodder.

Ranked, ready-to-write angles are driven by a prompt that already knows your audience, your content pillars, your hook structures, and your serial arcs before it ever goes out and looks at LinkedIn or Substack or Reddit.

This is the Claude Code skill I built to do that work. It’s called /research, and it produces a ranked, filtered, archetype-tagged angle list every Sunday before I touch a single essay.

Quick context: what’s a Claude Code skill?

If you’re already running Claude Code, skip this section.

If you’ve mostly used Claude in the web chat (or even dabbled in Cowork), here’s the short version:

• Claude Code is Anthropic’s terminal-based version of Claude. You work inside a folder on your computer. Claude reads your files, writes to them, and runs reusable workflows on demand.

• A skill is one of those reusable workflows. It’s a saved prompt that does one specific thing well, every time you call it.

• You trigger a skill by typing a slash command (/skill-name). It runs, reads the files it needs, and produces the output.

The /research skill runs once a week (usually Sunday) and replaces about three hours of manual scrolling, prompt iterating, and angle selection with roughly twelve minutes of structured scan.

The manual version is the tax

Before this skill existed, my Sunday research routine was:

1. Open ChatGPT or Claude. Ask what’s trending in AI tools, in Claude features, in solo operator content. Get back surface-level lists.

2. Open LinkedIn. Scroll the people I’ve benchmarked. Note what’s pulling traction.

3. Open Substack. Check the AI category for posts performing well. Note the angles.

4. Open Anthropic’s release notes. Check what shipped this week.

5. Try to map it all back to my pillars. Filter for what fits my content strategy. Discard the rest.

6. Try to map it back to my serial arcs. Does anything feed the series or themes I am focusing on?

7. Pick angles. Sit with the list. Wonder which one is worth writing first.

That was three to four hours every Sunday. Some Sundays the list ended up unranked, half-pillar-fit, and I’d write whichever angle had the strongest hook in my head at the time.

Reasoning that should have driven those decisions was buried somewhere in the scroll, but by the time I got to the writing I was too tired to surface it cleanly.

That’s the tax. It shows up in the slow drift from the strategy and in the quality of what eventually goes out the door. The danger to this slow drift is it be can absorbed for a long time before it becomes obvious.

What /research actually does

The skill reads four files when it runs:

1. Brand context. Who the audience is, what the content pillars are, which serial arcs are active, which frameworks I own.

2. Voice profile. So angles surface in the language I actually use.

3. Archetype reference. The six content types I publish into. Every angle gets tagged with the archetype it best fits.

4. Signature frameworks. So angles that connect to existing IP get flagged and prioritized.

Then it runs a structured scan against a defined source map.

The source map lives in the skill file, not in the prompt I type each week, which means every Sunday’s scan covers the same surfaces, in the same order, with the same depth.

No drift. No “oh I forgot to check Anthropic this week.”

Each topic it surfaces gets scored against six criteria:

1. Audience relevance. Does this matter to my readers right now?

2. Timeliness. Is this a live conversation this week, or a stale topic?

3. Hook potential. Can this become a 5–15 word hook using one of my five hook structures?

4. Framework fit. Does this connect to a framework I already own?

5. Pillar match. Which of my pillars does this serve?

6. Archetype fit. Which of the six archetypes does this fit best?

What comes back is a ranked list of angles already filtered for my audience. Each entry carries the metadata I need to make the writing call without re-doing any editorial work:

• Archetype

• Serial arc (if applicable)

• Hook type and suggested structure

• The signal: what’s being discussed and where

• The angle: my specific entry point

• Framework connection

• DM keyword suggestion

• Source URLs (for fact checking - don’t skip!)

Here’s what one of those entries looks like when it comes back. This is the angle that became the essay you’re reading right now:

ANGLE 4 of 11   |   Total score: 27/30
​
Archetype:       Skill Drop
Serial arc:      Claude Skills in the Wild
Hook structure:  Quantified Result
Suggested hook:  “Three hours of Sunday content research.
                  Twelve minutes with one Claude Code skill.”
​
Signal:          Three solo-operator Substacks this week wrote
                 about their content prompting routines, none of
                 them with a real fix. Anthropic’s skill
                 registration update on April 18 added a fresh
                 release-note hook for a Skill Drop angle.
​
Angle:           Walk through the /research skill end-to-end.
                 Show source map, scoring rubric, install
                 instructions. Lead magnet: redacted skill file.
​
Framework conn:  The Three Levels of AI Usage —
                 /research lives at Level 3.
​
DM keyword:      RESEARCH
​
Sources:         4 LinkedIn permalinks,
                 anthropic.com/news/claude-code-skills-update,
                 3 trending Substack URLs in the AI category

That’s the format for every angle that comes back. Each tag is a decision the angle has already made for me that I just need to screen and approve. By the time Monday morning hits, the editorial work is done.

The writing starts cold and immediate. I pick the top-scoring angle that fits the slot, open the essay file, and start drafting from a hook that’s already been pressure-tested against the rubric.

Subscribe now

What changes when the angle list is structured

The first scan I ran this way replaced about three hours of Sunday scrolling with one twelve-minute command and a ten-minute review of the output.

That’s the time math, but the bigger shift is what gets written.

When the angle list is unranked, I default to whichever hook felt strongest in my head Sunday night. Usually that means the news-reaction angle wins because news has the most pull in real time.

When the angle list is ranked and filtered, the themes or series I’ve created get fed every week without me having to remember to feed them. The skill explicitly checks each theme’s slot and flags any that came back empty before I commit to the week’s drafts.

This means I see pillar drift coming. The output names which series are about to starve, before I’ve spent any writing time on the imbalance.

Install instructions

If you want to build a version of this for your own brand:

1. Create a research.md skill file in your .claude/skills/folder (or wherever your Claude Code skills live).

2. Define your source map. Six to eight sources specific to your brand and the week your audience is actually living in.

   • The release notes for the tools your audience uses.

   • The publications your audience actuallyreads (not the ones it’s supposed to read).

   • The two or three platforms where the real conversations are happening this quarter.

   For a SaaS marketing operator that might be ProductHunt launches, your category competitors’ changelogs, the LinkedIn creators in your space, and three niche category Substacks.

   For a real estate broker it might be NAR releases, MLS market updates, three industry-specific Substacks, and the LinkedIn agents pulling traction in your market.

   Specificity here is what makes everything downstream possible.

   The scoring rubric can’t tell what matters from what doesn’t if the sources you’re scoring are noise to start with.

3. Write the scoring rubric. The six criteria above are a starting point. Adapt them to your brand. Audience relevance and pillar match are non-negotiable; the others should reflect what you actually use to make a “write this or skip it” call.

4. Reference your context files. The skill should read your pillars, your voice profile, your frameworks, and your archetypes (or whatever your equivalent shapes are). It can’t filter without knowing what to filter against, and it can’t tag an angle for archetype or framework fit if those things aren’t actually defined somewhere it can read.

5. Define the output format. A ranked block per angle with the metadata tags shown earlier. Format matters more than people expect. When Monday morning hits, you want each angle to land as a self-contained block you can read in ten seconds, structured so the only question that actually matters [which one am I writing?] is obvious at a glance.

Run it weekly. Iterate the source map and the scoring criteria as you see what surfaces.

Get the skill file

Want a starter skill file you can adapt to your brand to jumpstart this?

No hype. No theory.

AI is moving fast. I want to help you move with it.

I write this newsletter to close the gap between AI adoption and AI integration. Just systems and strategies I’m actually using in my business, with my clients, right now.

If this was helpful, send it to someone else who’s trying to figure out where they actually stand.

Share

Mine has tools I demoted, systems I scrapped, and beliefs I held onto longer than I should have. Every one of them felt right at the time, and every one got replaced by something more precise or better aligned.

The path to real growth is almost never linear. It’s iterative. Point A to point B? There’s usually a lot of detours and nuance in getting there.

Learning to cut losses or pivot is a (painful) skill that is sharpened through trial and error.

There’s been a lot of errors along the way, but here’s 7 missteps I made while refining my own processes and beliefs around working with AI that have stuck with me longer than the rest.

1. I picked an image tool early and stayed loyal too long

Dropped: A template-driven generator that fit when my visuals were simple, but couldn’t handle where my content was going.

Moved to: Gemini 2.5 Pro (with thinking) as primary for any complex visual work.

My visuals used to be simple: quote cards, single-subject headers, brand-consistent templates. The tool I used was built for exactly that (and did it well).

Then my goals changed. My personal brand and my business brands both started rewarding denser, more value-packed visuals. Infographics with multiple callouts, real hierarchy, and the kind of information density that respects an audience’s intelligence.

The template-first generator couldn’t hold that.

Gemini 2.5 Pro (with thinking) holds complexity the way a skilled designer would. Specificity in the prompt gets honored instead of smoothed away, and I can work at the level of detail my content requires.

The lesson isn’t about any specific generator because tools change all the time.

2. I tuned prompts when the leverage was in context files

Dropped: Prompt packs, prompt libraries, and the hours I used to spend polishing individual prompts for recurring tasks.

Moved to: Durable context files that every project and skill reads before producing anything.

Things like: a voice profile, anti-AI writing rules, defined content strategy, ICP definitions for every brand, signature frameworks, brand guidelines, master offer suite.

Quick breakdown for anyone newer to this layer of AI: A context file is a plain markdown document that tells an AI who you are, how you work, and what you refuse to produce (often more important) so its output shows up already aligned.

Mine capture how I write, who I write for, the frameworks I rely on, and the non-negotiables every draft has to pass. The AI reads them at the start of every session and shows up already trained instead of guessing.

The mistake wasn’t using prompts. It was believing the leverage lived in the prompt when it was always upstream.

This is the correction I wish I’d made sooner. Tuning prompts kills productivity and creativity.

3. I kept reaching for ChatGPT when Claude was the better workspace

Dropped: ChatGPT as the default AI for substantive work.

Moved to: Claude as the primary workspace.

Claude Code for production, long-running projects, and anything that needs to read from or write to my files. Cowork to systematically update and build. Chat for strategy work.

ChatGPT is fine for one-offs: a quick rewrite, a narrow lookup, a task that doesn’t need to remember anything about me or what I built yesterday.

When the work needs to persist or builds on frameworks I’ve already created, the comparison stops being about features and starts being about compounding context.

Claude holds my system. Claude Code in particular lets me work inside my own file structure, so everything I need is one reference away. It also runs the systems I’ve automated so I’m just running commands instead of mile long prompts.

Match the task to the right tool.

4. I ran everything as ad-hoc chats instead of building reusable skills

Dropped: Re-explaining the same task every time I needed it done.

Moved to: Named Claude Code skills that encode each recurring workflow and produce consistent output on command.

Quick context for anyone newer to Claude Code: A skill is a reusable workflow saved as a markdown file that you trigger by typing /skill-name.

It can include instructions, required inputs, quality checks, and references to your context files. Instead of opening a chat and re-specifying the whole task every time, I type /decompose and the skill handles every platform.

What used to take ninety minutes now takes fifteen.

Every ad-hoc chat was a chance for the workflow to drift - different outputs on the same task, voice inconsistency across a week of content, quality gates that depended on how I happened to frame the job that day.

The version of me from a year ago would have called this over-engineering. She was wrong.

Real over-engineering is doing the same work manually for the fiftieth time because building the reusable version felt slow (and intimidating) on day one.

5. I tried to invent systems from scratch instead of reverse-engineering what was already working

Dropped: Designing every template, framework, and workflow from a blank page as if originality were a virtue.

Moved to: Two parallel reverse-engineering practices that drive almost every system I build now.

1. Studying patterns already proven in the wider world. Operators and strategists whose work produces the outcomes I want to see in mine. Not copying them, but taking apart what makes their work land and rebuilding that mechanic inside my own voice and priorities.

2. Staying in my own lane and reverse-engineering what’s already worked for me. Content that performed, workflows that held up at volume, client engagements that produced the result, offers that closed. The patterns inside my own track record are often more useful than the ones I’m admiring in someone else’s work. They already account for my voice, my constraints, and my audience.

Every “original” system I tried to invent from zero took about ten times longer than one I built by studying something proven, whether the proof came from someone else’s work or from my own.

Originality is one of the most expensive habits a solo operator can have. The leverage isn’t in the invention. It’s in the adaptation.

The answer is almost always yes. Find the pattern, adapt it, keep moving.

Subscribe now

6. I let “I’m not a coder” define what I was allowed to build

Dropped: The belief that a non-technical background meant I couldn’t ship real software, and the self-imposed ceiling that came with it.

Moved to: Building skills, plugins, automation chains, and tool integrations I wouldn’t have dreamed of a year ago, with Claude Code as the pair programmer.

A year ago, I would have told you my lane was strategy, content, client work, and advisory. Software and terminal work belonged to someone else - someone who’d spent years inside a terminal (and even knew what that was) or who had an IT degree.

It required me to get specific about what I wanted and let Claude translate the specification into code. I describe the outcome, name the edge cases, and iterate when something breaks.

The code is almost never the bottleneck. The spec is. And writing a clear spec transfers directly from strategy work, content work, and every other kind of structured thinking you’ve already done.

What used to require a developer now requires a clear spec and a willingness to iterate (and maybe the bravery to step into a lane you hadn’t ever considered).

If you’re waiting until you “learn to code” before you build with AI, you’re waiting for a gate that’s already open. Start building.

7. I waited until I felt ready instead of shipping the messy version

Dropped: Researching, planning, and polishing until something felt safe to put in front of people.

Moved to: Shipping v1 into real conditions, letting use expose what needs to change, iterating in public.

Every week I waited was a week the system didn’t compound. Every draft I held back was data I didn’t get.

The feedback that changes your work doesn’t come from more research. It comes from use.

My version of perfectionism rarely looks like chasing the perfect draft. It looks like “one more round of prep” (aka anxiety wearing a project-plan costume).

I’ve fallen into it, and I still catch myself there. The correction is always the same: ship what I’ve got, let it meet the world, adjust from what I learn rather than from what I was afraid might happen.

Confidence follows execution. Not the reverse.

If I could go back and tell past-me one thing, it wouldn’t be “take it more seriously” or “prepare more.” It would be: ship, adjust, ship again.

What all seven have in common

Every one of these was only learnable by building the wrong version first.

That’s the part most productivity content skips. Reading someone else’s retrospective doesn’t always help you avoid the mistakes; it helps you notice yours faster. And noticing faster is the whole game.

The operators who compound are the ones willing to retire what they just built the moment it stops earning its place. Tools, systems, beliefs. None of it is precious.

Have you gotten caught in any of these traps?

No hype. No theory.

I write this newsletter to close the gap between AI adoption and AI integration. Just systems and strategies I’m actually using in my business, with my clients, right now.

AI is moving fast. I want to help you move with it.

Subscribe now

If this was helpful, send it to someone else who’s trying to figure out where they actually stand.

Share

About two-thirds of the class walked in believing they were using AI reasonably well. They’d copied prompts off Instagram. They’d used Claude or ChatGPT to help with a listing description.

On paper, they were adopters. When I asked them to put those same concepts into practice on a real task from their own work, the gap between what they thought they knew and what they could actually do showed up immediately.

The other third of the room was quieter. They were skeptical.

Some had tried AI once, got mediocre results and wrote it off. Some were worried about what it would mean for their work and if it could be trusted to run tasks where they had only ever depended on their own skills.

A few hadn’t really tried at all and weren’t sure they wanted to (fair, honestly).

By the end of the day, both groups were surprised by the same thing - the sheer capacity of what it could do in sequence and how it could be customized with context.

The sandbox opened dramatically (or for the first time).

“Using AI” isn’t one skill.

There are three levels of it right now, and naming them honestly is the most useful thing most professionals can do for themselves this year.

The Three Levels

Level 1: Chatting

You open a fresh conversation. You ask a question. You read the answer. You either use it, tweak it, or keep chatting until the output gets closer to what you wanted.

The ceiling at Level 1: Things you could have produced yourself, but faster.

Level 2: Prompting

You’ve figured out that how you ask matters as much as what you ask. You’ve built structured inputs that produce reliable, repeatable outputs. You have prompts you reuse. You give the model a role, a goal, constraints, and examples.

The ceiling at Level 2: Your best work, consistently, instead of occasionally.

Level 3: Building

You’ve stopped asking AI to do things and started designing systems where AI does things.

Multi-step workflows. Agents that take an input, run it across several tools, and return a finished output. Claude Code writing and debugging a codebase while you’re in another meeting. Cowork organizing your files, pulling three PDFs into a spreadsheet, and compiling research into a briefing document while you’re on a call.

The relationship stops being a conversation. It becomes a production line you direct.

The ceiling at Level 3: Work you could not have produced at all before, at volumes you could not have sustained.

These three levels are not close together. The World Economic Forum reports that workers with advanced AI skills earn 56% more than peers in the same roles.

My read of that number is that almost all of the premium lives at Level 3.

The Self-Diagnostic

Before we get to the markers: being at Level 1 is not a failure. Neither is being skeptical or just starting out. The median professional in April 2026 sits somewhere between Level 1 and Level 2 with occasional Level 3 moments. That isn’t a flaw in you. It’s a description of where most of the professional world currently lives.

AI can be intimidating, and the instinct to project a level of fluency you don’t quite have yet is completely human. This isn’t a ranking. It’s a way to see yourself clearly, because the version of you that can see the gap is already further along than the version that thought there was no gap.

Score yourself on your last ten AI sessions. Not the session you’re proud of. The actual last ten.

You’re probably at Level 1 if:

• You start most sessions from a blank conversation

• You edit AI output more than you edit your own first drafts

• Your prompts change from session to session, even for the same type of task

• You copy-paste AI output into other tools by hand

• You can’t point to a task AI now handles end-to-end without you in the middle

• Your wins with AI feel like lucky breakthroughs, not repeatable processes

You’re at Level 2 if:

• You have prompts you reuse, and they live somewhere retrievable (a doc, a Projects folder, a saved template)

• You get consistent output quality for the same task across different sessions

• You’ve stopped writing the same instructions from scratch every time

• You can explain why a prompt works, not just that it does

• You edit AI output to taste, not to fix fundamental errors

You’re at Level 3 if:

• There are tasks running on your behalf right now that you didn’t have to start manually

• You’ve connected AI to your files, your calendar, your data, or your codebase

• You think in workflows instead of conversations

• Your production volume has shifted in ways that would be physically impossible at Level 2

• You can point to work that exists because you built a system, not because you sat down and did it

If you landed closer to Level 1 than you expected, that isn’t a verdict on your intelligence or your career. It’s a starting point.

You climb from where you are, one rung at a time, and the first rung is owning which one you’re on right now.

What to Do This Week

Skip to the track that matches where you actually landed.

If you landed at Level 1: Build a prompt library

You don’t need a new tool. You need one file.

Step 1. Open a new note in whatever you already use. Apple Notes, Google Doc, Notion, a Claude Project. Name it My Prompts.

Step 2. For your next five AI sessions, when you write a prompt that actually works, paste it into that file with a one-line label. That’s it. No organizing, no tagging, no polishing.

Step 3. Before you start your next session on a similar task, open the file first. Reuse before you restart.

The habit you’re training into yourself is a simple one: before you hit enter, give the model a role, a goal, constraints, and one example of what good looks like. That single move closes most of the gap between Level 1 and Level 2.

Give it two weeks. You’ll have ten prompts you actually trust, and the version of yourself you thought you were will catch up to the version you actually are.

Once that muscle is built, if you want to see the full range of places AI could fit into your specific workflow, run this workflow audit prompt. It interviews you about how you actually work and hands back a map of where AI fits, what it could automate, and where to start first.

If you landed at Level 2: Run one Cowork task this week

You already know how to prompt. What you’re missing is the experience of watching AI finish a task on your machine without you being the middle step.

Step 1. Download the Claude desktop app from claude.ai/download if you don’t already have it. Cowork requires a Pro or Max plan ($20/month, or $17/month on annual).

Step 2. Audit your last ten AI sessions. Find one task where Claude gave you an output and you had to manually move it into another tool — a spreadsheet, an email, a doc, a folder. That’s your Level 3 candidate.

Step 3. In the desktop app, click the Cowork tab. Point it at a folder on your computer where the raw material for the task lives. Then describe the task the way you’d describe it to a capable colleague.

A good first task to try:

“Read the three PDFs in this folder. Compare them on [your criteria]. Produce a one-page summary I can save as a markdown file in this same folder.”

Don’t try to build a whole system on day one. Run one task. Feel the difference. Let the experience update your mental model before you scale anything.

When you’re ready for the full Cowork walkthrough — the folder structure, the first workflows to build, and the mistakes that waste the first week — I wrote the complete setup guide here.

If you landed at Level 3: Build your context folder

Your bottleneck isn’t tool fluency anymore. It’s context.

Step 1. Create a folder on your computer called something like “ai-context” or “[your name]-brain.” This is the folder Cowork and Claude Code will read for everything they need to know about you.

Step 2. Inside it, create five markdown files to start:

• about-me.md — who you are, what you do, what you care about, how you think

• voice.md — how you write, what you avoid, what “good” sounds like for you

• workflows.md — how your work actually flows, not the theoretical version

• clients.md (or audience.md) — the context about the people you work with that you’d otherwise explain every time

• offers.md — what you sell, who it’s for, which offers are live and which are in development

Step 3. Point Cowork at that folder (or keep it in a parent folder with your active projects). Every new task now starts with that context already loaded, so you stop explaining yourself from scratch every session.

Once the context is in place, build workflows against it.

One I use every week is an AI-powered meeting prep system that runs in 10 minutes — a good working example of what a Level 3 workflow actually looks like once the infrastructure is there to support it.

Why This Moment Matters

Two things happened in 2026 that you might not have connected yet.

One.

IEEE Spectrum, Fortune, and Fast Company all ran variations of the same headline in Q1:

“prompt engineering is dead.”

The standalone job title that paid $200K two years ago has been quietly retired.

Most people took that message as “don’t bother learning this.”

The actual signal was the opposite. Per PE Collective data, roles requiring prompt engineering skills (regardless of what the job was called) grew 3x between 2024 and 2026.

The title died because the skill stopped being a specialty and became a baseline. The same way “using Excel” stopped being a resume line and became an assumption.

Two.

Anthropic released Claude Cowork to general availability on Pro and Max plans this April. If that sentence didn’t mean much, here’s the plain version:

• Cowork is a mode inside the Claude desktop app.

• “General availability” means it’s open to anyone on a Pro or Max plan (starting at $20/month). No waitlist. No developer credentials.

• What it does: It takes actions on your computer. Reads your files. Builds spreadsheets. Pulls research from multiple websites into a briefing document. Organizes a folder. Drafts and saves outputs across multiple tools in one run. Not as suggestions you copy-paste. As finished work on your machine.

• Why it matters: Until this month, this kind of work required Claude Code, terminal commands, and a level of comfort with technical troubleshooting most professionals reasonably didn’t want to build. That barrier just fell.

The combination is the point. Prompt engineering got absorbed into the baseline at the same moment agentic AI workflows became accessible without code. The floor rose and the ceiling opened in the same window.

That’s why the gap between Level 2 and Level 3 is the one that actually matters right now.

The gap between Level 1 and Level 2 is narrowing on its own as the tools get smarter.

The Level 2 to Level 3 gap is a different kind of gap. It’s the one between asking and designing, and that shift is cognitive, not technical.

The prerequisite used to be “learn to code.” As of this month, it’s “change your mental model.”

If you want the longer read on why this shift is happening and what it means for your role specifically, I wrote about it here.

Subscribe now

One Principle That Makes the Difference

One of the frameworks I use with clients is called the AI Infrastructure Integration Model, and its core principle is this: governance precedes scale.

You can’t scale AI usage past a certain point without building infrastructure underneath it, or you produce chaos faster than you produce value.

That principle applies at the individual level just as cleanly. The context folder in the Level 3 track is the individual-scale version of governance. Build it before you scale your usage, and every level above it holds.

Level 3 without infrastructure collapses in a week.
Level 3 with infrastructure compounds indefinitely.

TL;DR

The three levels:

• Level 1 (Chatting): Asking questions, reading answers, copy-pasting output. Ceiling: faster versions of work you already do.

• Level 2 (Prompting): Reusable, structured inputs that produce consistent outputs. Ceiling: your best work, consistently.

• Level 3 (Building): Multi-step workflows and agentic tasks you direct. Ceiling: work you couldn’t have produced at all before.

The uncomfortable read: Most professionals who think they’re at Level 2 are behaviorally at Level 1. The gap that matters in 2026 isn’t 1→2. It’s 2→3.

What changed this month: Claude Cowork went generally available on Pro and Max plans. The Level 3 building layer is now accessible without code. The prerequisite used to be “learn to code.” Now it’s “change your mental model.”

Action steps by level:

1. Level 1: Start one My Prompts file this week. Reuse before you restart for the next five sessions.

2. Level 2: Audit your copy-paste moments. Run one through Cowork this week using the three-PDF prompt above. One task, not a system.

3. Level 3: Build the five-file context folder (about-me, voice, workflows, clients, offers) and point Cowork at it.

Wednesday’s essay walks through exactly what Level 3 looks like in my own workflow, including the specific tasks I’d recommend trying first if you want to feel what it actually feels like.

No hype. No theory.

I write this newsletter to close the gap between AI adoption and AI integration. Just systems and strategies I’m actually using in my business, with my clients, right now.

AI is moving fast. I want to help you move with it.

If this was helpful, send it to someone else who’s trying to figure out where they actually stand.

Share

I don’t mean the elevator pitch version.

I mean the real version: every step, every tool, every decision point, every workaround you’ve built over the years that lives entirely in your head.

There’s so much nuance, it’s hard to articulate.

Recently, I helped a friend of mine, a home appraiser with over 15 years of experience, map his entire workflow.

He knows his job cold. But when I asked him to walk through every step between receiving an order and delivering a finished report, the gaps were immediate.

He’d say “I just look up comps,” which turned out to be a 45-minute process involving three different systems, manual data entry, and a spreadsheet he’s been maintaining since 2019.

That experience became the foundation for a prompt I now use with every workflow I audit, mine included.

How This Works

You’re going to paste a detailed prompt into Claude that turns it into an operational interviewer.

It will ask you 100 questions across 10 categories, covering your professional context, how work enters your system, how you gather information, how you analyze and produce deliverables, how you handle quality control, and where your bottlenecks live.

Important: answer honestly, not theoretically.

The value of this exercise comes from documenting how you literally work, not how you think you should work.

Include the workarounds hanging on by hope and good intentions, the duct-tape spreadsheets, the steps you do on autopilot that you’ve never written down and don’t even think about anymore.

That’s where the automation gold is.

At the end, Claude compiles everything into a diagnostic report with 9 sections, from a full workflow map to a prioritized 90-day implementation plan.

This doesn’t just stop at pinning down your workflow.

It takes it to the next level with how to optimize it.

Subscribe now

This post includes affiliate links to tools I actually use. I may earn a small commission if you purchase through them — at no extra cost to you.

Before You Start: Three Tips

Tip 1: Use Claude for this.

The prompt is designed for Claude’s interview style: one question at a time, follow-up depth, structured output, and it works in the free tier.

If you’re running this in the Claude desktop app, even better — you can save the output directly to your files.

1. Go to claude.ai/download. Download the app.

2. You can do this on the free account, but I highly recommend a Pro account ($20/month, or $17/month if you pay annually).

3. Open the app. Click the Chat tab at the top.

Tip 2: Save the output as a .md file.

Markdown (.md) files are just plain text with simple formatting.

They’re readable everywhere and they’re the format AI works with best if you want to reference this document in future conversations, in Cowork, or in another AI tool like ChatGPT.

How to create a .md file if you’ve never done it:

The easiest way is to create a Google Doc, write or paste your content, then go to File → Download → Markdown (.md).

You can also use a dedicated markdown editor like Typora, which is what I use daily for a cleaner editing and reading experience.

Tip 3: Dictate your answers instead of typing them.

This is 100 questions. That’s a lot of typing.

When people type, they edit themselves. They clean up their language, polish their phrasing, and inadvertently describe the theoretical version of how they work instead of the real version.

When you talk, you don’t do that.

You describe what actually happens, the 3 shortcuts you actually use, the “I don’t know why I do it this way but I do“ steps, the informal processes you’ve never thought to write down.

That conversational honesty is exactly what this interview is designed to capture.

I personally use Wispr Flow for this. It’s a voice-to-text tool that runs in the background on your computer.

You talk, it types.

For a deep interview like this where you’re describing complex processes, dictating is tremendously faster than typing, and the output is more authentic because you’re not self-editing.

The Prompt (Copy Everything Below)

Paste this entire prompt into Claude. Replace the bracketed sections with your own role/industry, then let it run.

You are an AI Workflow Architect and Diagnostic Interviewer specializing in identifying automation, augmentation, and future-proofing opportunities for knowledge workers.

Your job is to conduct a deep operational interview with a [YOUR ROLE / INDUSTRY] in order to:

1. Fully understand their current workflow

2. Identify time-consuming manual processes

3. Identify tasks AI could automate or accelerate

4. Identify tasks that should remain human-driven

5. Identify strategic risks from AI disruption

6. Identify tools and systems that could increase output volume and efficiency

Your goal is to create a complete operational blueprint of the workflow and AI opportunities.

You are not here to give quick suggestions. You are here to reverse engineer the entire workflow first.

---

INTERVIEW PHILOSOPHY

Most professionals cannot clearly explain how they actually work. They describe theoretical workflows, not what actually happens day-to-day. Your job is to uncover the truth of their workflow.

You must:

• Push past vague answers

• Ask for concrete examples

• Clarify tools used

• Identify time spent on tasks

• Identify bottlenecks

• Identify repetitive processes

Your goal is to understand every step between receiving work and delivering the final output.

---

INTERVIEW STRUCTURE

Conduct 100 questions total across the following categories. Do not follow the categories rigidly if the conversation reveals useful threads — pursue them.

SECTION 1 — PROFESSIONAL CONTEXT (10 questions)
Understand their professional environment: years of experience, role scope, industry, workload, team structure, compensation model, technology familiarity, current constraints.

SECTION 2 — WORK INTAKE & JOB MANAGEMENT (10 questions)
How work arrives and gets organized: sources of assignments/projects, management systems, acceptance criteria, scheduling, client communication, administrative workload.

SECTION 3 — INFORMATION & DATA GATHERING (15 questions)
How they collect the information needed to do their work: research methods, systems used, data sources, verification steps, manual data entry, time spent on research per project.

SECTION 4 — CORE TASK EXECUTION (10 questions)
The primary “doing” portion of their role: what happens during execution, tools used, notes/documentation created, how work product gets built, handoffs between steps.

SECTION 5 — ANALYSIS & DECISION-MAKING (15 questions)
The most analytical portion: how they evaluate, compare, assess, and make professional judgments. Methods, frameworks, manual vs. tool-driven analysis, time spent.

SECTION 6 — DELIVERABLE CREATION (15 questions)
How final outputs are produced: software used, templates, narrative/written components, repetitive sections, revision cycles, formatting time.

SECTION 7 — QUALITY CONTROL & REVIEW (10 questions)
Review before delivery: self-review process, compliance checks, common errors, revision requests, time spent correcting.

SECTION 8 — ADMINISTRATIVE TASKS (5 questions)
Non-core work: emails, scheduling, file organization, invoicing, delivery logistics.

SECTION 9 — BOTTLENECKS & TIME DRAINS (5 questions)
Explicitly identify inefficiencies: tasks that feel repetitive, unnecessary but required, create delays, or require excessive manual work.

SECTION 10 — FUTURE RISK & STRATEGY (5 questions)
Their perspective on AI disruption in their field: concerns, tasks requiring human expertise, skills needed, competitive positioning.

---

INTERVIEW RULES

1. ONE QUESTION AT A TIME. Never ask multiple questions at once. Ask one. Wait for the response. Continue.

2. PUSH FOR SPECIFICITY. If they give a vague answer like “I just research it,” follow up: What tool? What steps? How long does it take?

3. ASK FOR EXAMPLES. “Walk me through the last project you completed.” “How long did that step take?” “What tool did you use?”

4. IDENTIFY TIME SPENT. When possible: “Roughly how many minutes does this step take per project?”

5. FOLLOW INTERESTING THREADS. If an inefficiency appears, explore it further before moving on.

6. CHALLENGE ASSUMPTIONS. If they say something must be manual, ask why.

---

FINAL OUTPUT

After all 100 questions are completed, produce a comprehensive diagnostic report with the following sections:

1. PROFESSIONAL PROFILE — Summary of role, workload, and environment.

2. CURRENT WORKFLOW MAP — Step-by-step reconstruction of the full workflow from work received to final delivery. Include every step, tools used, and estimated time per step.

3. TASK BREAKDOWN BY CATEGORY — Tasks organized into: data gathering, core execution, analysis, deliverable creation, admin, quality control.

4. TIME ALLOCATION ANALYSIS — Estimated time per category. Identify largest time drains.

5. AUTOMATION OPPORTUNITIES — For each workflow step: tasks AI could fully automate, tasks AI could partially assist, tasks that must remain human. Include recommended tools where applicable.

6. TOP 10 HIGHEST-ROI IMPROVEMENTS — For each: what process it improves, how AI would be used, estimated time saved per project.

7. AI TOOL STACK RECOMMENDATIONS — Specific categories of tools for research, analysis, content generation, document review, scheduling, and administration.

8. FUTURE-PROOFING STRATEGY — Tasks most at risk of AI disruption, tasks where human judgment remains essential, skills that increase long-term value.

9. 90-DAY IMPLEMENTATION PLAN — Month 1: quick efficiency wins. Month 2: workflow upgrades. Month 3: advanced AI integration.

---

Begin by asking the first question from the Professional Context section. Remember: one question at a time. Push for specific answers. Continue until all 100 questions are completed.

What You’ll Get Back

After the interview, Claude compiles a diagnostic report that covers:

• A complete map of how you actually work (not the theoretical version, the real one)

• Every step categorized by time investment and automation potential

• A clear separation of what requires your judgment versus what’s repetitive execution

• The top 10 highest-ROI places to integrate AI, with estimated time savings

• Tool recommendations specific to your role

• A 90-day plan so you’re not trying to change everything at once

This isn’t a generic “tips for using AI” article.

This is a structured operational audit built from your own answers about your own work; that’s how you help future-proof yourself.

Best Practices

I originally built this prompt for that appraiser friend.

We ran it, and the diagnostic revealed meaningful areas where his workflows could be automated to enhance the parts of his role that require human judgment, which directly increased the volume he could handle per day.

Since then I’ve adapted it for clients, colleagues, and team members. The sections flex to fit any knowledge-work role because the interview is built around your answers, not a pre-built template.

A few things I’d recommend:

Block an evening or take time this weekend.

The interview is thorough. That’s the point. Surface-level audits produce surface-level results.

Knock this out once, and it’s the foundation to an infinite number of workflows and tasks. Time spent here will pay dividends.

Answer like you’re talking to a colleague, not writing a job description.

Include the workarounds, the messy parts, the “I’ve been doing it this way since 2020 for no good reason” steps. That’s where the value hides.

Save the diagnostic report.

This becomes a reference document you’ll use repeatedly: for AI integration planning, for delegation, for onboarding, for pricing decisions.

Save it as a .md file in a place you can find it. If you use Claude’s Cowork mode, drop it in your Cowork folder so Claude has it as context for future tasks (more on that in my Cowork setup guide).

TL;DR

Open Claude. Paste the prompt from this post. Replace the bracketed role/industry with yours. Block out the time and answer honestly.

You’ll walk away with a document, built from your own words, that shows you exactly how you work, where your time goes, and what a realistic AI integration plan looks like for your specific role.

No $500 course. No generic tutorial. Just an honest conversation about your work that produces something you can actually use.

Subscribe now

No hype. No theory.

I write this newsletter to close the gap between AI adoption and AI integration. Just systems and strategies that I’m actually using in my business, with my clients, right now.

AI is moving fast. I want to help you move with it.

If this was helpful, send it to someone else who’s figuring out where AI fits in their workflow.

Share

Someone sent this your way? Thanks for being here! Subscribe for free, and stay in the loop on what’s working today.

Subscribe now

Over coffee last month, he told me AI is being integrated broadly across their operations at a pace that would have been unthinkable two years ago. In the next six months, their workflows are going to look drastically different from where they are today.

But here’s what caught my attention: the company chose to adopt AI. That was the easy decision. The hard part is a multi-generational workforce literally integrating it into how they work.

The 60-year-old analyst who’s been modeling manually for 35 years isn’t touching it (and is highly skeptical).

The 28-year-old built three experimental workflows in two weeks.

The 19-year old intern can’t imagine why anyone would not use ChatGPT everyday and drafted their resume with it.

The people in the middle are mostly passive. They’ll use it if someone hands them a prompt, but they aren’t reaching for it independently.

The gap isn’t adoption. It’s integration.

The people closing that gap fastest are doing something very specific.

The Simplest Starting Point

The people actually pulling ahead right now aren’t learning AI as a separate skill.

They’re taking work they already do - focusing stuff they’re already good at - and finding the parts where AI handles the execution so they can focus on the aspects that require their brain, their judgement.

That distinction matters. The conversation right now is full of people telling you to “learn AI.” That’s vague to the point of being useless.

The useful version is 3 simple steps:

1. Look at your existing workflows

2. Identify which steps are repetitive execution

3. Start testing whether AI can handle those steps while you direct the outcome

That’s not a career pivot. That’s a workflow upgrade. And it’s the single most practical thing you can do this week.

Subscribe now

This post includes affiliate links to tools I actually use. I may earn a small commission if you purchase through them — at no extra cost to you.

How to Actually Do This

Here’s a framework I’ve walked clients through and that I use personally for identifying where AI fits into work you’re already doing.

Step 1: Pick one workflow you repeat at least weekly.

Not your most complex project.

Something you do regularly that follows a recognizable pattern — a weekly report, a client update, a data pull and summary, a recurring email sequence, an internal status document.

The criteria: you’ve done it enough times that you could explain the steps to someone else.

Step 2: Document every step — honestly.

Not the idealized version, but the real version.

The one where you’re copying and pasting between apps, reformatting the same data you reformatted last week, and writing the same opening paragraph for the third time this month.

Most people discover that 40-60% of their “high-value” work is actually execution they repeat with minor variations.

That’s not a judgment on your work. That’s the nature of professional workflows. The question is whether you want to keep spending your time there.

If you want to go deep on this step, I built a prompt that turns Claude into a diagnostic interviewer: it asks you 100 targeted questions about your workflow, pushes past vague answers, and compiles the whole thing into a structured automation blueprint.

I published the full copy-paste prompt here: The AI Workflow Audit.

Knock this out in a day, and it’ll produce a document you’ll reference for months.

If you’re going to do one thing from this article, do that.

Step 3: Label each step — “requires my judgment” or “execution I repeat.”

This is the critical sort.

“Requires my judgment“ means the outcome changes based on your expertise, context, or professional discretion.

Deciding which data points matter. Interpreting what the numbers mean for this specific client. Choosing the right recommendation based on factors the data doesn’t capture.

“Execution I repeat“ means the step follows a pattern you could describe to someone else.

Pulling the data. Formatting the report. Drafting the summary. Structuring the email.

Be honest during this step.

Most people initially label everything as “requires judgment” and then realize, when they actually look at it, that a significant portion is patterned execution wearing a judgment costume.

Step 4: Test AI on one “execution” step.

Just one. Take the clearest, most repetitive step and run it through Claude with specific instructions: what the task is, what format you want, what to include, what to avoid.

Then evaluate the output against what you would have produced yourself. Where is it close? Where is it off? What would you change?

That gap — between what AI produces and what the final version needs to be — is your judgment in action. That’s the part that’s yours.

Step 5: Refine and repeat.

If the output was 70% usable, adjust your instructions and try again. If it was 90% usable, you just found 90% of a workflow step you can reclaim.

Then move to the next execution step.

Subscribe now

What This Actually Gives You

When you do this, even with one workflow, three things happen.

You get time back.

Not theoretical time. Actual hours.

The execution steps that used to eat your afternoons now take minutes, which means you have space for the work that actually moves the needle.

You get clarity on what you’re actually good at.

When the execution is handled, you’re left with the judgment calls, the strategic decisions, the creative thinking, the relationship management — the stuff that’s genuinely hard to replicate.

That clarity tells you exactly where your value lives.

You become the person who can articulate where humans are essential in your workflow and where they’re not.

That matters when roles get restructured, teams get reorganized, and the “what should we automate?” conversation comes to your department.

The person who already has the answer shapes the restructure. The person who doesn’t gets shaped by it.

TL;DR

Pick one workflow. Write down the steps. Sort them. Test one step with AI.

That’s 30 minutes. You don’t need a course, a certification, or a new job title. You need 30 minutes and one honest look at how you spend your time.

The people moving fastest right now aren’t doing anything complicated.

They’re just not waiting for someone else to figure out where AI fits into their work. They’re figuring it out themselves.

Subscribe now

No hype. No theory.

I write this newsletter to close the gap between AI adoption and AI integration. Just systems and strategies that I’m actually using in my business, with my clients, right now.

AI is moving fast. I want to help you move with it.

If this was helpful, send it to someone else who’s figuring out where AI fits in their workflow.

Share

Someone sent this your way? Thanks for being here! Subscribe for free, and stay in the loop on what’s working today.

Subscribe now

LinkedIn stalking. Googling their company and skimming their site. Reading their last few social posts. Scanning for recent news.

Trying to piece together enough context that I wouldn’t walk into the conversation cold.

It was necessary work, but it followed the same pattern every single time. Different person, same research steps.

That combination makes it one of the best places to put AI to work.

Now I do it in 10 minutes, and the output is more thorough than what I used to produce manually.

This post includes affiliate links to tools I actually use. I may earn a small commission if you purchase through them — at no extra cost to you.

The System: Three Parts

This research system has three components. Once you build it, running it on any new prospect takes about 10 minutes total.

Subscribe now

Part 1: The Research Prompt (Copy This)

This is the core of the system.

Paste the prompt below into your preferred AI search tool (I like Claude or Grok but you can use this with any of them - Perplexity, ChatGPT, Gemini, etc.) whenever you need to prep for a meeting.

Replace the brackets with your specifics.

If you have the time, pasting this into more than one AI tool and comparing the results will help you nail down the best tool for your particular workflow.

You are a Strategic Research Analyst preparing me for a professional meeting. I need a comprehensive, actionable briefing document on this person and their organization.

Who I’m meeting:
[Name, title, company]

Context for the meeting:
[What this meeting is about: discovery call, follow-up, partnership discussion, pitch, etc. Include any prior context you have, how you connected, what they’ve expressed interest in, anything relevant from previous conversations.]

My role/business:
[Brief description of what you do and what you’d potentially offer or discuss. This helps Claude tailor the briefing to what’s relevant for YOUR conversation, not a generic company overview.]

---

Research and produce a briefing document with the following sections:

1. PERSON PROFILE

• Professional background and career trajectory

• Current role and likely responsibilities

• Areas of expertise and professional focus

• Notable career moves, publications, or public positions

• Communication style indicators (based on their public content: do they post on LinkedIn? What tone? What topics do they engage with?)

2. ORGANIZATION OVERVIEW

• What the company does (specific, not boilerplate)

• Size, stage, and market position

• Recent news, announcements, or strategic moves (last 6 months with sources)

• Key challenges their industry or company size typically faces

• Competitive landscape context (who are they competing with, what pressures exist)

3. STRATEGIC CONTEXT

• What’s likely keeping this person up at night professionally (based on their role, industry, and current market conditions)

• Potential pain points or priorities relevant to what I offer

• What success looks like for someone in their position

• Triggers that might have prompted this meeting (industry shifts, company changes, growth stage)

4. CONVERSATION STRATEGY

• 5 specific questions I should ask (not generic, tailored to their situation and what I need to learn)

• Topics to lead with based on likely shared interests or relevant angles

• Topics to avoid or handle carefully

• How to position what I do in language that maps to their priorities (not my pitch, their frame)

5. QUICK REFERENCE CARD

• 3-sentence summary of who they are and what they care about

• The single most important thing to remember walking into this meeting

• One specific detail to reference that shows I did my homework (not a generic compliment, something specific from their content, background, or company activity)

---

Format: Scannable. Use headers, bold key points, and keep each section to 3-5 bullet points maximum. I’ll be reviewing this on my phone 10 minutes before the meeting. Make it fast to scan.

Tone: Direct, analytical, useful. No filler, no generic business platitudes. Specific and actionable.

Subscribe now

Part 2: Add Context Layers

The prompt above will get you 70-80% of the way there with just a name and company. But you can push the output quality significantly higher by adding context.

Before you run the prompt, spend five minutes gathering:

Their LinkedIn “About” section.

Copy-paste it into the prompt under the person’s name. This gives Claude their self-described positioning, which is gold for conversation strategy.

Their last 3-5 LinkedIn posts or articles (if they post).

Copy the text or screenshot it, even just headlines and brief summaries. This reveals what they’re currently thinking about, which topics they engage with, and how they communicate.

Any prior conversation notes you have.

If you’ve talked to them before, include what was discussed, what they said they were interested in, and any follow-up items. Meeting transcripts work well here too, if you have them.

Claude will weave this into the briefing so it doesn’t feel like a cold restart.

This takes five minutes because you aren’t doing the deep dive here. The output goes from “solid generic briefing” to “this person clearly did their research.”

Part 3: Build a Reusable Reference File

If you run this prompt regularly for weekly prospect calls, client meetings, or partnership conversations, build a reference file that contains your standard context so you don’t retype it every time.

Create a simple .md file (markdown, just a text file with basic formatting). Call it something like: meeting-prep-context.md.

Include:

• A description of what you do, who you work with, and what you typically offer

• Your standard discovery questions (the ones you ask in every first conversation)

• Common objections you encounter and how you typically address them

• Your ideal client profile: who’s a great fit and who isn’t (more important to define)

• Any relevant case studies or examples you reference frequently

How to create a .md file:

Open a Google Doc, write your content, then go to File → Download → Markdown (.md).

Or use a markdown editor like Typora, which is what I use for all my working documents. Clean interface, handles .md natively, and makes formatting effortless.

Once this file exists, you can paste or reference its contents whenever you run the research prompt.

But if you want to skip the copy-paste step entirely, there’s a better option.

The game changer: Claude has a feature called Cowork mode in its desktop app.

Instead of starting every conversation from scratch, you point Claude Cowork at a folder on your computer. Claude reads every file in that folder and uses all of it as context for whatever you ask it to do.

So if your meeting-prep-context.md file lives in that folder, Claude already knows your business, your discovery questions, your ideal client profile, and your positioning before you type a single word.

Take it a step further: Set up a “Proposals” sub-folder in your Cowork folder.

You can then ask Claude to reference that prospect’s file (name-company.md) before working on a task, and update the file with any new or relevant information to keep it fresh.

I wrote a setup guide for Cowork with step-by-step instructions here.

But if you want to get started right now:

1. Go to claude.ai/download. Download the app.

2. You need a Pro account ($20/month, or $17/month if you pay annually).

3. Open the app. Click the Cowork tab at the top.

4. Select a folder from your computer. This is how Claude reads your files.

5. Pro tip: create markdown files about you, or anything you want Claude to know.

6. Drop your meeting-prep-context.md file in that folder and you’re set.

What the Output Actually Gives You

When you run this prompt with a real prospect, the briefing Claude produces covers ground you wouldn’t get from 30 minutes of manual research.

For example: say you’re prepping for a discovery call with a regional property management firm.

You paste in the owner’s name and company, add their LinkedIn summary and a couple of recent posts.

Claude comes back with a briefing that includes their career trajectory, what the company actually does (specific, not the boilerplate from their website), recent news about expansion or operational changes, and five questions tailored to their situation.

Not generic discovery questions. Questions informed by what they’ve been posting about, what their company size suggests they’re dealing with, and where your work connects to their priorities.

The conversation strategy section is where this pays off most. Claude maps your positioning to their language.

Instead of walking in with your pitch, you walk in with their frame.

I review these briefings on my phone 10 minutes before calls. The difference between showing up with real context versus surface-level Googling is immediately obvious in how the conversation flows.

Make It Faster: Voice Input

One more tip that makes this system even faster.

If you use Wispr Flow (voice-to-text), you can dictate the context portion of the prompt instead of typing it.

Open Claude, paste the research prompt template, then use Wispr to talk through the meeting context:

“I’m meeting with Sarah, she’s the VP of Ops at XYZ, we connected through LinkedIn, she mentioned she’s interested in how we handle workflow automation for mid-market companies...”

Wispr transcribes it. Claude runs the research.

Total time from start to briefing document: under 10 minutes.

I use this for every meeting, and it has made a dramatic difference in expediting tasks.

TL;DR

1. Copy the research prompt above

2. Open Claude and paste it in

3. Fill in the brackets with your next meeting’s details

4. Add any LinkedIn content or prior notes you have (5 minutes)

5. Review the briefing 8-10 minutes before the meeting

Then walk into that conversation better prepared than 95% of the people your prospect talks to this week.

If you want to take this further, build the reusable reference file and drop it into a Claude Cowork folder so your meeting prep context is always available. I walk through the full Cowork setup here.

No hype. No theory.

I write this newsletter to close the gap between AI adoption and AI integration. Just systems and strategies that I’m actually using in my business, with my clients, right now.

AI is moving fast. I want to help you move with it.

If this was helpful, send it to someone else who’s figuring out where AI fits in their workflow.

Share

Someone sent this your way? Thanks for being here! Subscribe for free, and stay in the loop on what’s working today.

Subscribe now

Normally that takes me about two hours. Research, outline, draft, revise, format.

I did it in 10 minutes using Claude’s Cowork mode.

And it was more specific than what I’d typically produce under time pressure, because Claude already had everything it needed about my business before I even started the task.

I’m going to walk you through the entire setup from scratch, step by step, so you can build this system yourself today.

If you’ve never used Cowork (or never heard of it), that’s fine.

By the end of this article you’ll have it running.

What Cowork Is (and Why It’s a Game Changer)

If you’ve used Claude or ChatGPT in a regular chat window, you know the biggest limitation: you start from zero every time.

Every conversation, you’re re-explaining who you are, what you do, and what you want. Frustrating, right?

If you’ve been opening a chat window, typing a prompt, and getting output that sounds like it could’ve been written by anyone for anyone, this is the reason.

The model isn’t the problem. The missing context is.

Cowork changes that.

It’s a feature in the Claude desktop app where you point Claude at a folder on your computer. Claude reads every file in that folder (your documents, your templates, your notes, your examples) and uses all of it as context for whatever you ask it to do.

The more context you give Claude as files, the less prompting you need. The output goes from “generic AI slop” to “this actually sounds like someone who knows my business.“

The key shift: The work you do upfront building your folder is what makes every future task faster and better.

You’re not just prompting. You’re building an operational system.

Subscribe now

This post includes affiliate links to tools I actually use. I may earn a small commission if you purchase through them — at no extra cost to you.

Step 1: Download Claude Desktop and Open Cowork

1. Go to claude.ai/download. Download the app.

2. You need a Pro account ($20/month, or $17/month if you pay annually).

3. Open the app. You’ll see three modes across the top: Chat, Code, and Cowork. Click Cowork.

Important settings:

• Select Opus 4.6 as your model (it’s in the model dropdown). This is the most capable model and it makes a real difference in output quality for complex tasks.

• Turn on Extended Thinking (in the settings). This lets Claude reason through more complex tasks before responding.

Create your folder structure

Create a new folder somewhere easy to find.

I keep mine on my Desktop and call it “Claude Cowork.” You can name it whatever makes sense to you.

Add these sub-folders:

• ABOUT

• PROPOSALS [or “projects” or whatever task you are working on]

• CLAUDE OUTPUTS

• TEMPLATES [or “examples”]

• PROSPECTS [this is optional but helpful]

Step 2: Build Your Context Files

This is the step that makes everything work.

It’s also the step most people skip, which is why they think AI “doesn’t work well enough.”

Putting in the front-end sweat equity here pays dividends later.

You’re going to create a few .md files and drop them into your Cowork folders. These are just plain text files with simple formatting. Claude reads them automatically every time you start a task.

Markdown (.md) files aren’t just for Cowork. If you upload documents into ChatGPT, Gemini, or any other AI tool, .md files consistently produce better results than PDFs or Word docs because the AI can read the formatting structure cleanly.

Build these once and they work across every platform.

How to create .md files

If you’ve never made a markdown file, here’s the easiest way:

1. Google Docs method (easiest): Open a new Google Doc. Write or paste your content.

   Go to File → Download → Markdown (.md). Save it to your Cowork folder.

2. Markdown editor method (recommend for editing): I use Typora for all of my working documents. It’s a clean interface, handles .md natively, shows you formatted text as you type instead of raw code.

   Open Typora, write or paste your content, save directly to your Cowork folder as a .md file.

3. Quick method (any platform): Open any text editor (TextEdit on Mac, Notepad on Windows). Write or paste your content. Save the file and change the extension from .txt to .md.

   On Mac, make sure “All Files” is selected in the format dropdown so it doesn’t append .txt automatically.

The 4 files you need

File 1: about-me.md

Everything Claude needs to know about your business. Write this like you’re briefing a new employee on day one.

Include:

• What you do and who you serve

• How you typically structure engagements or projects

• Your pricing (tiers, ranges, what’s included at each level)

• Your positioning (what makes your approach different)

• Your professional background (relevant experience that informs your work)

Length: 1-2 pages is plenty. Be specific, not comprehensive.

File 2: voice-and-style.md

How you write and communicate.

This is the file that makes outputs sound like you instead of like a robot.

Include:

• How you describe your tone (professional but conversational? Direct? Warm? Technical?)

• Words and phrases you actually use

• Words and phrases you never want to see in your content (mine has a long list)

• 2-3 examples of writing you’re happy with (past emails, proposals, social posts, anything that sounds like you at your best)

• Any formatting preferences (paragraph length, use of headers, how you structure documents)

• All the statements, phrases, words you hate or would never say. This is arguably the most important part of the file (mine has an even longer list here)

This file is the difference between “AI-sounding output” and “output that actually sounds like me.” Invest the time here to be thorough.

File 3: proposal-template.md (or the deliverable you produce most)

Your preferred structure for the deliverable you want to speed up.

For this example, it’s proposals. For you, it might be client reports, project briefs, pitch emails, or status updates.

Include:

• Your preferred section order

• What each section should accomplish

• Approximate length per section

• 2-3 examples of past deliverables you were happy with (anonymize client details if needed)

Claude uses this as a structural blueprint. Instead of guessing how you organize your proposals, it follows your actual format.

File 4: anti-patterns.md

Everything you want Claude to avoid so it doesn’t sound like AI generated garbage. This is your quality filter.

Include:

• Generic corporate phrases you hate (”synergy,” “circle back,” “move the needle,” whatever makes you cringe)

• Structural patterns to avoid (overly formal tone, bullet-point-heavy formatting, generic openings)

• Common AI writing tics (if you’ve noticed patterns in AI output that don’t sound like you, list them here)

• Anything else that would make you rewrite the entire output

You do this once. Every task you run through Cowork from this point forward uses all of this context automatically.

That initial time investment pays dividends on every single deliverable you produce through this system.

Step 3: Use Claude to Help You Build The Files

If you’re staring at a blank page thinking “I don’t even know where to start with an about-me file,” you’re not alone.

Most people know their business inside out but struggle to articulate it in a structured document.

Here’s the shortcut: use Claude to interview you.

Open a regular Claude chat (you don’t need Cowork for this part) and try something like:

“I’m building a set of context files so you can help me with my work. Interview me about my business. Ask me one question at a time. Cover what I do, who I serve, how I structure my work, my pricing, and what makes my approach different. When we’re done, compile my answers into a clean about-me.md file.”

Claude will walk you through it conversationally. You talk (or type), it organizes.

You can do the same thing for your voice file.

Paste in 2-3 examples of writing you’re happy with and ask Claude to analyze your tone, sentence patterns, and word choices. Then tell it what you’d add or change.

The result becomes your voice-and-style.md.

This turns what feels like a daunting writing exercise into a 15-minute conversation. And if you use Wispr Flow to dictate your answers, it’s even faster. That’s what I do (goodbye carpal tunnel).

Step 4: Set Your Global Instructions

Global Instructions are the guardrails you give Claude before it touches anything in your folder.

Without them, Claude has full read-and-write access to every file in your Cowork directory.

That means it could theoretically overwrite your source files, delete things, create files in the wrong folder, or go browsing through folders that have nothing to do with the task you asked for.

That’s not a flaw. It’s just how the tool works by default. Global Instructions are how you fix it.

Go to Settings → Cowork → Scroll down to Global Instructions

Copy and paste the template below, then adjust the folder names to match your setup:

GLOBAL INSTRUCTIONS

Folder rules:

ABOUT/ → Read only. Never modify, rename, or delete any file in this folder. These are source-of-truth documents.

TEMPLATES/ → Read only. Study these for structure and patterns. Never modify templates.

CLAUDE OUTPUTS/ → Write here. This is the only folder where you create new files. All deliverables go here.

All other folders → Read only unless I explicitly tell you to edit something.

If you are unsure whether to read or write — ask me first.

Operating rules:

Never delete any file in any folder under any circumstances. If something needs to be replaced, create a new version in CLAUDE OUTPUTS with a clear name.

Never modify original source files. If I ask you to edit something, create a copy in CLAUDE OUTPUTS with changes applied.

Always confirm before creating files. Tell me the filename and where it will go before writing it.

Stay scoped to the current task. Don’t go browsing folders unrelated to what I’ve asked you to do.

Before every task:

Read about-me.md and voice-and-style.md first. No task starts without these.

If the task involves a deliverable, read the relevant template file before proceeding.

If you’re unsure about tone or positioning, re-read the voice file before proceeding.

That’s it. Modify it as you go. Add rules when Claude does something you don’t want, remove ones that feel unnecessary. The file evolves with your workflow.

This takes five minutes to set up and prevents the moment where Claude overwrites your carefully built voice file with a “helpful” revision you didn’t ask for.

Step 5: Run the Task (The 10-Minute Part)

With your folder built, here’s what producing a proposal actually looks like:

Add your prospect intel

Before you ask Claude to write anything, add context about this specific prospect.

You can either drop a quick note into the folder or include it directly in your message to Claude.

I usually include:

• Company overview (a paragraph copied from their website works fine)

• The specific problem they mentioned on our call. I keep brief notes after every discovery conversation. You can also include meeting transcriptions if you use something like Granola.

• Relevant details about budget, timeline, or scope

• What they said they’re looking for

You don’t need to format this. Just the raw information. Claude will organize it.

If these are prospects you are likely to work with in the future or do follow up outreach to, I recommend setting up the optional “Prospects” sub-folder I mentioned earlier in your Cowork folder.

You can then ask Claude to reference that prospect’s file (name-company.md) before working on a task, and update the file with any new or relevant information to keep it fresh.

Give the task

In the Cowork tab, it’ll ask you to select a folder on your computer. This is where all your context files live, and where Claude will save anything it creates for you.

Then, type something like:

“Write a custom proposal for [Prospect Name / Company] using the prospect intel I’ve provided. Follow the structure in my proposal template. Match my voice and style. Avoid everything in my anti-patterns file. Save the output as a .md file.”

That’s the entire prompt. Your context files are doing the heavy lifting.

Hit enter. Claude reads your context files first, processes the prospect intel, and generates a full proposal saved directly to your folder.

Review and edit (5-7 minutes)

Open the file Claude created.

In my experience, if your context files are thorough, the first draft is at least 80% usable immediately:

• Structure will match your template exactly

• Tone will be close to your natural voice (because of your voice file)

• Prospect details will be woven in naturally

• Pricing will be accurate to your tiers

What I typically change:

Personal touches Claude can’t know.

“When you mentioned X on our call, that stuck with me because...” These are the details that make a proposal feel human. I usually add 2-3 of these.

Judgment calls on emphasis.

Sometimes Claude weights sections evenly when I want to lead heavy on one area because I know what this particular prospect cares about most. That’s a quick rewrite.

The occasional tone miss.

Despite the voice file, a phrase slips through sometimes that doesn’t sound like me. Quick find-and-replace.

Most of that is adding the human layer, not fixing errors.

A task that used to take me a couple of hours now takes about 10 minutes.

After you finish that one-time setup (building your context files), the math gets interesting fast.

Even at one proposal per week, you’re reclaiming over 90 hours a year. At three per week, it’s closer to 300.

That time goes somewhere else. For me, it goes toward the parts of the work that actually require me.

P.S. You can use these files and this set up to draft any number of deliverables: content, website copy, digital products, course outlines, etc.

And now, they sound like you. Without you having to repeat yourself a dozen times.

That’s the difference between a framework that gets you 30% of the way there on the first draft versus one that gets you 80%+ of the way there in the first go.

Going Deeper: Use This With Your Workflow

If you’ve run the AI Workflow Audit and have a diagnostic of your workflows and exactly where AI fits in them (or should), drop that document into your Cowork folder too.

If you do nothing else from this article, do that. The audit will give you the roadmap for all the places AI can automate steps or entire tasks in the next 90 days.

Once that’s done, Claude now has:

• Who you are and what you do (about-me.md)

• How you communicate (voice-and-style.md)

• How your deliverables are structured (template files)

• What to avoid (anti-patterns.md)

• A complete map of your workflows and where AI fits (your audit document)

That’s an operational context layer that makes almost any task you throw at Claude faster and more accurate.

The folder compounds: every file you add makes future outputs better.

TL;DR

1. Download Claude desktop: claude.ai/download

2. Click Cowork. Create your folder. Select Opus 4.6 and turn on Extended Thinking.

3. Build your four context files (about-me, voice-and-style, template, anti-patterns). Use Claude to interview you if you’re not sure where to start.

4. Set up your Global Instructions

5. Add prospect intel for one real deliverable

6. (Optional but recommend) Include your Workflow Audit for deeper context

7. Run the task. Time it. Compare.

Then ask yourself what you’d do with those extra hours every week if you got them back.

That’s the question worth answering.

Subscribe now

No hype. No theory.

I write this newsletter to close the gap between AI adoption and AI integration. Just systems and strategies that I’m actually using in my business, with my clients, right now.

AI is moving fast. I want to help you move with it.

If this was helpful, send it to someone else who’s figuring out where AI fits in their workflow.

Share

Someone sent this your way? Thanks for being here! Subscribe for free, and stay in the loop on what’s working today.

Subscribe now