extract-lessons
Skill39 starsChanged 22 days agoMITView on GitHub
Harvest a finished session for anything that would help beyond this one project, decide honestly where each piece belongs, and put it there. Most sessions yield nothing, and saying so is the correct outcome. Separates a project detail from a general trap, routes real findings to the right skill as a pull request, spots when a cluster of homeless lessons means a new skill is needed, and recognises when a body of knowledge is its own domain and belongs in a separate repo rather than bloating this one. Use at the end of a session, when the user asks whether anything is worth pulling into The Last Mile or another toolkit, or says "anything useful from this", "extract the lessons", "add this to the toolkit".
- Installs packages
In this repository21 other files it gives its agents.
- .claude/CLAUDE.mdCLAUDE.md
- build-a-baselineSkill
- connect-toolsSkill
- fact-checkSkill
- feature-demoSkill
- improve-skillSkill
- lean-outSkill
- mine-communitySkill
- one-time-surfaceSkill
- port-featureSkill
- reference-huntSkill
- researchSkill
- screen-atlasSkill
- search-pagesSkill
- see-it-workSkill
- session-endSkill
- set-goalsSkill
- setupSkill
- shipSkill
- store-listingSkill
- syncSkill
The file
We split this skill into the rules it gives an agent
12 rules. About 0 are common, 9 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈9
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 31
If no, it belongs in the project's own rules file (CLAUDE.md or AGENTS.md) and nowhere else.
- Distinctive: line 53
Anything still contested: If the session ended with two sources disagreeing and no measurement settling it, it is not knowledge yet.
- Distinctive: line 56
Anything that only makes sense with your numbers in it: If stripping the project's metrics leaves nothing, there was nothing.
- Distinctive: line 77
Before writing a new skill, ask what the cluster is actually about.
- Distinctive: line 93
If it IS a skill here, write it the way the others are written: a short description that says when to use it, numbered steps, tables over …
- Distinctive: line 102
Every metric: revenue, user counts, view counts, conversion rates
- Very specific: line 111
Add one line under the unreleased heading in CHANGELOG.md.
- Distinctive: line 112
Add the skill to the table in README.md if it is a new one.
- Very specific: line 113
Run node scripts/validate.mjs and fix what fails.
- Very specific: line 115
Commit only what you changed, push the branch, open a pull request following .github/pull_request_template.md.
- Distinctive: line 118
If you cannot push to the repo, fork it and open the pull request from there.
- Distinctive: line 129
If a cluster suggested a new repo, say so as a question, not a decision.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Extract lessons
Anyone running several projects wants the toolkit to get better from each one. This turns a finished session into whatever small, general thing it actually produced, and most often that is nothing.
This is not improve-skill. That one fixes a skill that misbehaved. This one
looks for knowledge the toolkit does not have yet.
The rule that makes this worth running
The default outcome is nothing, and that is a success. A skill that finds something every time fills a public repo with one project's anecdotes, and the next person reads them as advice. Expect to extract from maybe one session in four. If you find yourself reaching, stop.
1. List the candidates
From the session, write down every thing learned, in one line each. Include the mistakes, which are usually the best material. Do not filter yet.
Then, for each one, answer the only question that matters first:
Would this still be true for an app that is not this one?
If no, it belongs in the project's own rules file (CLAUDE.md or AGENTS.md)
and nowhere else. Write it there and strike it off. This removes most of the
list, and removing it is the point.
2. Apply the evidence bar
What is left has to survive a second test, and it is stricter than it looks.
Extract these:
| Kind | Why it survives | Example shape |
| A platform fact | Checkable by anyone, today | "A photo post takes a title capped at 90 characters AND a separate description. A video takes one caption" |
| A trap, with its mechanism | The reader can recognise it before it costs them | "Fine on a desktop preview and wrong on the phone means the platform is drawing its own UI over your image, not cropping it" |
| Arithmetic or a method | Produces the reader's own answer, not yours | "Never pay a cost per thousand views above your own revenue per thousand views. Compute it as revenue over installs, times installs per thousand views" |
| An exact invocation that works | Saves an hour of the next person's life | "yt-dlp -U fails for a pip install. Use <that python> -m pip install -U yt-dlp" |
Do not extract these:
- "This worked for us." Six posts, one campaign, one account. A result is not a method, and a stranger acting on your anecdote is the harm this bar exists to prevent.
- Anything still contested. If the session ended with two sources disagreeing and no measurement settling it, it is not knowledge yet. Leave it in the project's research notes.
- Anything that only makes sense with your numbers in it. If stripping the project's metrics leaves nothing, there was nothing.
A useful phrasing test: could you write it as a trap or a formula with no reference to your app at all? If not, it fails.
3. Find each survivor a home
| The lesson is... | Goes to |
| A correction to how an existing skill behaves | improve-skill. Stop, run that instead. |
| New knowledge an existing skill should carry | That skill, as an added step, trap or table row. |
| Orphaned, and the only one of its kind | A one line note in the relevant skill's references, or nowhere. Do not build a skill for one lesson. |
| Orphaned, and part of a cluster of three or more related orphans | A new skill. See step 4. |
When several lessons have nowhere to live, that is the finding. It means the toolkit has a gap, not that the lessons are weak. Say so plainly rather than forcing them into the nearest skill that half fits.
4. A new skill, or a different repo
Before writing a new skill, ask what the cluster is actually about.
This repo is about getting an app shipped and found. A cluster inside that subject becomes a skill here.
A cluster that is its own domain does not. Content marketing, hiring, fundraising, customer support: each is a field with its own practice, and absorbing one into a shipping toolkit makes both worse. The honest move is to say: "this is three skills' worth of a different subject, it wants its own repo." Then stop and let the user decide. Never create a repo unprompted.
Two signals that it is a separate domain rather than a skill here:
- Somebody could want this toolkit and not want that one at all.
- Explaining the cluster needs vocabulary this repo's other skills never use.
If it IS a skill here, write it the way the others are written: a short description that says when to use it, numbered steps, tables over prose, and the honest limits stated in the skill rather than discovered by the user.
5. Strip the project out
The repo is public. Before anything is committed, remove:
- The app's name, the user's name, handles, URLs and repo paths
- Every metric: revenue, user counts, view counts, conversion rates
- Screenshots, file paths and command output that identify the project
- Anything a competitor would find interesting
The lesson has to stand with all of that gone. If it does not, it was step 2's "only makes sense with your numbers in it" and should not have reached here.
6. Open the pull request
- Add one line under the unreleased heading in
CHANGELOG.md. - Add the skill to the table in
README.mdif it is a new one. - Run
node scripts/validate.mjsand fix what fails. It checks, among other things, that no Markdown file contains a long dash. - Commit only what you changed, push the branch, open a pull request following
.github/pull_request_template.md.
If you cannot push to the repo, fork it and open the pull request from there.
7. Tell the user
A few lines, and be specific about what you threw away:
- What you extracted, and where it went.
- What you considered and rejected, and why. This is the most useful half. "Four things looked like lessons and were one project's results" tells the user their session was normal, not empty.
- The pull request link.
- If a cluster suggested a new repo, say so as a question, not a decision.
Then note that the project picks the change up on its next sync, and that the
synced copy under .claude/skills/ must never be edited by hand.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

