self-learning

Self-learning: harvest golden paths into skills

This skill turns something you just figured out the hard way into a reusable Agent Skill, so the next session — yours or a teammate's — starts already knowing the proven route instead of rediscovering it from scratch.

It is a meta-skill: it doesn't do the work, it captures how work got done. It's tool-neutral — it works with any agent that reads the Agent Skills format (e.g. Claude Code and Codex, which both load SKILL.md skills natively). Where a step differs by tool, the generic version comes first and any tool-specific detail is only an example.

Recognize the moment

Watch for these signals during normal work. Any one of them is a cue to harvest:

• A task only worked after several attempts, wrong turns, or a correction from the user. The successful path is worth more than the failures around it.
• You discovered project-specific facts the agent didn't know up front: where creds/env vars live, which selector or backend talks to a service, a non-obvious command, a required sequence, a gotcha that defies the obvious assumption.
• It's an operational workflow likely to recur: reach the dev/prod DB, deploy, run migrations, seed data, verify a change live, run one specific test path, rotate a key, tail the right logs.
• The user signals it explicitly: "remember this", "save this as a skill", "don't make me re-explain this next time".

Act on the cue immediately — don't ask for permission first, whether the user requested it or you noticed it yourself. Harvest the skill, then tell the user what you captured and where (step 5). They can always edit or delete it.

Skill, memory, or skip?

Not every lesson deserves a whole skill — triage first, so you don't bloat the skills list with one-liners:

• A multi-step, reusable procedure or workflow (how to deploy, reach the DB, run the migration dance, verify live) → harvest it as a skill using the procedure below.
• A single standalone fact or one-line correction (an env var name, a path, one gotcha) → if your harness has a lightweight memory/notes facility (e.g. a MEMORY.md index), record it there instead; a whole skill is overkill for a one-liner. With no such facility, make a small skill.
• A genuinely one-off thing unlikely to recur → skip it.

When you do harvest, capture the failures too, not just the win: the approaches you ruled out and why often save more time next session than the golden path itself.

Promotion rule: don't enshrine guesses

A skill is authoritative — the next session trusts it without re-deriving it — so hold promotion to a high bar. Only write a skill when all three hold:

1. A passing check. The path was actually verified — a test passed, the command exited clean, the repro reproduced, the build went green. Record what the check was. "Seemed to work" is not a passing check.
2. A named failure pattern. You can name the failure this path avoids or diagnoses (e.g. "stale build cache → phantom type errors"), not a vague "sometimes it breaks".
3. At least one ruled-out dead-end. A concrete approach you tried and eliminated, with the reason.

If any is missing, it isn't a skill yet — leave a tentative note in memory (marked unverified) or skip it. This keeps confident guesses out of the skill set.

Harvest procedure

• 1. Apply the promotion rule (above). Passing check + named failure pattern + one ruled-out dead-end — or it isn't a skill: note it in memory or skip. Don't proceed on a confident guess.
• 2. Choose scope and name yourself using the heuristics below — don't stop to ask. Default to project scope; pick a clear, specific name.
• 3. Dedupe. Look for an existing skill to UPDATE rather than duplicate. List your agent's skills directories — the project one and the user-level one (e.g. Claude Code .claude/skills + ~/.claude/skills, Codex .codex/skills + ~/.codex/skills, or your tool's equivalent). Also glance at any memory/notes index — a fact already there may just need a pointer.
• 4. Distill the golden path from THIS conversation before delegating — while it's fresh in your head: the exact working commands, file paths, env var names, the required order, and (just as important) the dead-ends to avoid. This is the raw material for the write.
• 5. Delegate the write to a subagent that inherits this conversation if your tool supports one, or do it inline otherwise — see below. The conversation is the only place the golden path lives, so whoever writes it must have that context.
• 6. When the write is done, relay the new skill's path to the user and, in one line, what it captured.

Scope: project vs global

• Project (the repo's skills directory — e.g. .claude/skills/, .codex/skills/): the path is specific to THIS codebase — its env vars, its build/release steps, its schema, its quirks. Most harvested operational skills are project-scoped, and they ship to the team via git.
• Global (your user-level skills directory — e.g. ~/.claude/skills/, ~/.codex/skills/): the path generalizes across projects — a personal tool, a cross-repo habit, or a workflow tied to your machine rather than to one repo.

When unsure, prefer project — an over-shared global skill triggers in repos where its commands don't apply.

Delegate the write (subagent, or inline)

Whoever writes the skill needs THIS conversation's context — it's the only place the golden path lives. Two equally valid ways to run it:

• Inline — do the steps yourself in the main loop. Always works.
• Subagent — if your tool can delegate to a subagent that inherits this conversation, use it to keep the harvesting work out of your main context. (Claude Code: a skill with context: fork. Codex and others spawn subagents their own way.) Don't hand it to a fresh agent with no context — it would start blank with nothing to extract.

Either way it over-reaches by default, so box it in tightly. Follow this brief (fill in the bracketed parts) — hand it to the subagent, or work through it yourself inline:

You are harvesting a skill. Your ONLY job is to write a new Agent Skill capturing the golden path we just worked out in this conversation: [one-line description of the workflow].

Hard rules:

• Write ONLY under [skills dir]/[skill-name]/. Do NOT modify project source, run builds, install anything, or resume the original task.
• First read [this-skill-dir]/references/skill-authoring.md and [this-skill-dir]/assets/SKILL.template.md, then author SKILL.md to that spec, plus any references/ or assets/ files the procedure warrants.
• Capture the PROCEDURE — commands, paths, the required order, gotchas — not a one-off answer. Generalize so it works next time.
• Capture the FAILURES too: the approaches we ruled out and why, so the next session skips the dead-ends. Put them in a "What didn't work" section.
• Enforce the promotion rule: the skill must record the passing check that verified this path, name the failure pattern it addresses, and list at least one ruled-out dead-end. If any is missing (e.g. nothing was actually verified), STOP and report it isn't promotable — leave a tentative memory note instead of writing the skill.
• NEVER write secret VALUES (passwords, tokens, connection strings, API keys). Record only WHERE to find them: the env var name, the selector function, the MCP tool, the secret manager. Reproducing a secret into a skill file leaks it.
• Self-validate before finishing (see the checklist in skill-authoring.md).
• Report back: the absolute path you wrote and a one-line summary. Then STOP — do not pick the original task back up.

Gotchas

• Secrets never go in a skill file. Skills get committed and open-sourced. Point to where the secret lives; never reproduce the value. This is the single most important rule in this skill.
• Run this 10-second self-audit the moment before you call write_file — paste the proposed body through this filter in your head:
  • Does any line contain a literal Bearer , Basic , sk-, ghp_, xox[bpars]-, AKIA, or a 32+ char base64/hex blob? → redact it.
  • Is there a connection string (postgres://, mysql://, mongodb://, redis://, https://user:pass@, amqp://, kafka://)? → strip credentials, keep scheme + host + path.
  • Did the user paste a value earlier in the conversation that you're now re-typing? → replace with the source pointer (env:FOO_TOKEN, ~/.aws/credentials, MCP tool secrets/get, etc.).
  • Is the value a placeholder (YOUR_TOKEN_HERE, <api-key>, ${TO_FILL}) that will be filled by a hook? → still redact; placeholders that get auto-filled leak the same way literals do.
  • Is this an example URL in a docs snippet? → use example.com, user:[email protected], or a documented test fixture.
• name must equal the directory name, and be lowercase a-z/0-9/hyphens only — no leading, trailing, or doubled hyphens. A mismatch means the skill won't load.
• Whoever writes the skill over-reaches by default (a subagent especially). That's why the brief above forbids touching project source or resuming the task — keep it boxed to the skills directory.
• Don't duplicate. If a near-identical skill (or memory) already exists, update it instead of spawning a second one that competes to trigger.
• Capture procedures, not answers. "Join orders to customers for EMEA" is useless next time; "how to find the right tables and build the query" is the skill. See references/skill-authoring.md.
• Keep SKILL.md tight (< 500 lines, < ~5000 tokens). Push detail into references/ and tell the reader when to load each file.

For the full authoring spec, see references/skill-authoring.md. The fill-in template is assets/SKILL.template.md.