This morning I sat down, opened a folder on my computer, and typed "read _READ-ME-FIRST.md and pick up where we left off." My AI read the file and briefed me: what we finished yesterday, what files it made, what's next. Then we got to work.
That's my whole system now. It took about a year to get here.
For most of that year my "system" was a pile of prompts in a note and a Custom GPT I kept re-explaining myself to. It worked, but I rebuilt the same setup every time I started something new, and my best workflows lived in my head where I couldn't reuse them. What changed was small: I stopped keeping my work inside a chatbot and moved it into a folder my AI works in.
You're past the prompting basics. Most AI content is either too basic or too dev-flavored. This is the missing middle: turning one working setup into something you reuse and hand off, without writing a line of code.
But before you build anything, audit. Building the wrong system well is the expensive mistake here.
Step 0: audit first (do this before you build)
Most "deeper workflow" energy gets spent automating something that wasn't worth automating. Ten minutes here saves a week.
Act as an operations auditor for my AI use.
I'll paste a list of the tasks
I use AI for in a normal week.
For each, rank by: time it costs me, how
often it repeats, and how standardized the inputs are.
Then tell me which ONE is the best candidate
to turn into a reusable system,
and which ones I should NOT systematize yet and why.
Push back if I'm about to over-build.If you can't write down what "done" looks like, you're not ready to template it yet.
🧘The template: 3 parts you clone per use case
Once the audit tells you what's worth building, create these three pieces.
Part 1: context block (150 to 300 words)
Remember the context profile you saved as a file at the end of the last move? That's where this starts. Trim it down for this specific system: who it's for, what you'll give it, what you expect back, the rules it should always follow. One focused context per system is better than one giant profile used for everything.
Part 2: worked examples (2 to 3)
This is what makes the system reliable. Paste 2 to 3 real before/after pairs: a messy input you actually had, and the output you'd call correct.
The model copies the pattern from examples far more reliably than from instructions. Examples are the spec.
Part 3: rules file
A short plain-text file the system follows every time. Don't type it from scratch. Grab the ready-to-fill rules-file template (attached below), drop in your brackets, and you've got the first one in a few minutes.
# [System name]
## What this does
One line.
## Inputs
What you'll paste in.
## Rules
- Do X. Never do Y.
- When unsure, ask, don't guess.
- Flag anything unverified before presenting it.
## Output format
Exactly what the result should look like.
## Example
One worked input -> correct output.Keep each rules file single-responsibility: one file, one job. Five to ten focused files that don't overlap are better than one sprawling mega-prompt.
Where I keep mine: a Cowork folder
Here's the part that changed how I work. I stopped keeping these files inside a chatbot and moved them into a folder on my computer that my AI works in. The app is called Cowork: it lets Claude read and edit the files in that folder.
I won't teach Cowork here, you'll set it up your own way. I just want to show you why it's my home for all of this now:
No more copy-pasting. I point Claude at a file in the folder instead of pasting context into every chat. The files are on my computer, so I can edit one by hand, tell Claude what I changed (or show it a screenshot of the edit), and ask it to review my work. That back-and-forth is the part I love.
The system remembers itself. This is the move. I keep one file, _READ-ME-FIRST.md, and Claude updates it at the end of each session: what we finished, what files it created and where, and what's next. When I sit down again I just say "read _READ-ME-FIRST.md and pick up where we left off." The system briefs me instead of the other way around.
[SCREENSHOT TO CAPTURE (no attachment exists for this one yet). Take a shot of my real Cowork folder: _READ-ME-FIRST.md sitting at the top, the project folders beside it. Caption: this is the actual folder I built this whole launch in.]
No Cowork? The same three files work in a Project. Pick the setup that fits your tools.
Why this works
Examples carry more than instructions, and a folder Claude operates in beats a chat it forgets. Keep each system to one job, and when one breaks you know exactly which one. Clone it per client or per content type and you've got something you could hand to a teammate.
Where it breaks
Multi-agent before you need it is a trap. If a single well-specified system does the job, don't wire up a 5-agent pipeline that fails silently in five places.
Stale rules slowly make the output worse. A file you set in March makes April's answers subtly wrong, without telling you. Version it.
And the auditor prompt will always find a reason to build something, because that's the question you asked it. You make the final call, not the model.
One honest limit on Cowork: it's a paid desktop app, not a phone tool, and long sessions burn tokens. Keep your context files lean or it spends its budget reading you instead of working for you.
Your 20-minute win
Run Step 0 on your work this week. Take the single task it ranks highest and build Part 1 and Part 2 for it: a context block and two real examples. Add the rules file tomorrow.
Run it today, not "when things calm down." They won't.
How was today's issue?
If this felt like a lot
You don't need it yet. The honest sequence is the context-profile move first. One profile, one home. This is what it grows into once that's a habit. Nothing here expires. It'll be waiting.
Build one template this week, audit and all. Next week I'm opening a room for people working at this level, where a tested template you post becomes a shared asset other builders clone, with your name on it.
Dan Rice · AI Signal Read once. Use AI better all week.