agentleFS
sblueminsbluemin/fleet-harness/.claude/skills/learning-harvest/SKILL.md

learning-harvest

Skill18 starsChanged yesterdayMITView on GitHub

Distill proven recurring friction from substantive work into reusable rules, skills, or tests, or prune stale instructions. Not for one-off incidents, speculation, or restating code/git history.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Keep non-obvious learning from completed work in one retrievable, actionable home.

    line 8
  2. Distinctive:

    Check existing homes for duplicates first.

    line 34
  3. Distinctive:

    Verify links, commands, and ownership.

    line 42
  4. Distinctive:

    If nothing survives, say so briefly and finish.

    line 44

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Learning Harvest

Keep non-obvious learning from completed work in one retrievable, actionable home. Do not let instructions accumulate faster than signal. This skill does not add product features.

Trigger and approval

On explicit request or after substantive work, briefly propose candidates. Do not claim an automatic hook ran. New persistent encodings and deletions require user approval of the candidates. Do not repeat approval for a specific update already authorized by the request.

Candidate gate

Review actual friction, discoveries, judgment calls, and skills used. A candidate must satisfy all three:

  • Recurrence: will the situation happen again?
  • Cost: was the obstacle costly or non-obvious enough?
  • Generality: does the learning apply beyond that single incident?

Discard candidates that fail any condition. Incorrect commands or missing skill boundaries are candidates too, but one model mistake does not justify a universal model rule.

Route to one home

LearningPreferred owner
Reproducible regression/defect conditionTest or automated check
Repeatable procedureExisting skill; new skill only if no suitable home exists
Procedure-specific trapThat skill's conditional reference/gotcha
Stable, expensive-to-violate rule needed before scoped workNearest CLAUDE.md
User preference or working feedbackPersistent memory

Check existing homes for duplicates first. Propose migration rather than copying when a better home exists. CLAUDE.md additions must pass root risk-weighted minimalism; do not accumulate handbooks or inventories. Prefer enforceable tests/scripts over prose when feasible.

Propose, apply, verify

  1. Encode each survivor concisely as symptom → action → why (failure mode). Include evidence for all three conditions, the owning home, and duplicates/migration.
  2. Inspect the touched surface and propose keep/update/delete dispositions too. Pruning review is required; unnecessary deletion is not.
  3. Apply only approved items. Repository changes use a dedicated worktree; Memory follows its current write contract. For approved repository changes, follow the root CLAUDE.md standing delivery authorization through pr-workflow, honoring local-only requests and narrower scopes; this does not approve an unapproved candidate or replace memory approval.
  4. Skill edits use narrow descriptions, conditionally loaded references, and explicit completion/stop conditions. Keep stable boundaries rather than adding model-performance claims or ceremonial investigation/testing steps.
  5. Verify links, commands, and ownership. Check that the rule resolves its target case without blocking neighboring normal cases. Do not persist unverified generalizations.

If nothing survives, say so briefly and finish. Stop at proposal when awaiting approval; after authorized encoding, report locations, changes/removals, and verification. Do not prolong the task with endless retrospection or new-skill creation just to manufacture learning.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.