Skill Harvest

A library of skills only compounds if what you learn gets back into it. The default failure is silent: you solve something non-obvious, the session ends, and the lesson dies with the context window. The next agent — maybe you tomorrow — re-learns it from scratch. This skill is the reflex that closes the loop: notice the lesson, decide whether it's worth keeping, and route the keepers into the library through [[skill-creator]].

Harvest is the trigger and triage layer. It does not author skills — [[skill-creator]] does that. Harvest decides whether, what, and where; skill-creator handles how.

When to Use

Run a harvest pass when the session produced a reusable lesson, especially:

• A debug that took real effort and the root cause was non-obvious ([[fault-recovery]] just ran)
• You (or the agent) made the same class of mistake you've made before — a pattern worth a guardrail
• A workflow or sequence that finally worked after false starts, that you'd want to repeat
• A skill failed to fire when it should have, fired wrongly, or was ignored under pressure
• A skill's steps were incomplete or out of date versus what actually worked
• A recurring task type has no skill and you improvised the same process twice ([[skill-router]] §7)
• A review, incident, or migration surfaced a rule that isn't written down anywhere

Skip when:

• The work was routine and an existing skill already covered it well — nothing new was learned
• The lesson is a one-off fact about this codebase (that's a repo doc / [[decision-docs]], not a skill)
• You're mid-task — harvest at a natural boundary, not as a distraction from finishing

Not a substitute for [[skill-creator]] (the authoring craft) or [[decision-docs]] (project-specific records). Harvest points to those; it doesn't replace them.

Process

1. Scan the session for harvest candidates

Before ending the work, ask explicitly: what did I learn here that a future agent would want? Look for the signals above. Write each candidate as one sentence — the lesson, not the incident.

Weak:   "We fixed the login bug."
Strong: "OAuth state cookies silently drop when SameSite=Lax and the IdP posts back cross-site —
         the fix is SameSite=None; Secure. Worth a guardrail in hardening."

If nothing clears the bar, say so and stop — a forced harvest produces noise that dilutes routing.

2. Triage each candidate — what kind of thing is it?

Most weak harvests come from skipping this step and writing a "skill" for something that's really a doc or a one-off. Triage honestly: a skill encodes a repeatable workflow where corners get cut.

3. Check it doesn't already exist

Search the library before creating anything (grep -ri "<keyword>" skills/ agents/). Overlap dilutes routing worse than a missing skill. If a skill nearly covers it, improve or link that one rather than spawning a near-duplicate.

4. Decide new-skill vs improve-existing — bias toward improving

• Improve existing when the lesson sharpens, corrects, or extends a skill that already fires for

this work. Cheaper, keeps the library tight. Most harvests land here.

• New skill only when the work is a distinct, recurring job no existing skill owns, and you can

name the trigger conditions an agent would see.

5. Hand off to skill-creator

Pass the triaged, deduped lesson to [[skill-creator]] with the destination decided. Skill-creator owns the anatomy — trigger-rich description, executable Process, Common Rationalizations, Red Flags, checkbox Verification — and the validation/catalog steps. Don't reinvent that here.

For a guardrail or router fix, make the edit directly in [[agent-guardrails]] / [[skill-router]] and still run validation.

6. Validate and catalog

Run scripts/validate.sh and scripts/generate-catalog.sh (and scripts/skill-test.sh if you added a behavioral test). A harvest that breaks structure or skips the catalog isn't done ([[pipeline-ops]] enforces this in CI).

7. Confirm the loop closed

State what was harvested and where it went — or that nothing met the bar. A harvest with no decision and no diff is a harvest that didn't happen.

Common Rationalizations

• "I'll remember this." — You won't, and the next agent never had it. The context window is not the library.
• "It's not worth writing up." — If it cost you real time to learn, it'll cost the next person the same; the bar is reusable, not novel.
• "Just make it a new skill." — Most lessons sharpen an existing skill; a near-duplicate dilutes routing more than a gap does. Triage first (step 2).
• "Harvest = write a skill." — Harvest is triage; the lesson might be a guardrail, a router fix, a doc, or nothing. Deciding that is the work.
• "I'll harvest everything to be safe." — Forced harvests produce noise; an unfocused library is worse than a smaller sharp one.
• "I'm out of context, skip it." — The moment you have the lesson is the only moment it's cheap to capture; later it's gone.

Red Flags

• A non-trivial debug or repeated mistake ends with no harvest pass at all
• A "new skill" that's really a one-off fact about this repo (belongs in [[decision-docs]])
• A harvested skill that overlaps an existing one instead of improving/linking it
• Triage skipped — every lesson defaulted to "new skill"
• Lesson written as the incident ("we fixed X") rather than the transferable rule
• Handed to [[skill-creator]] without checking for an existing skill first
• Validation / catalog not run after the change
• "I'll harvest later" on a lesson you only hold in the current context

Verification

• [ ] Session scanned for candidates; each written as a transferable lesson, not an incident
• [ ] Each candidate triaged to a destination (new / improve / guardrail / router / doc / drop)
• [ ] Searched the library; no near-duplicate created — existing skill improved or linked when one fit
• [ ] New-vs-improve decided with bias toward improving an existing skill
• [ ] Keepers handed to [[skill-creator]] (or guardrail/router edited directly), not authored ad hoc here
• [ ] validate.sh + generate-catalog.sh run (and skill-test.sh if a test was added)
• [ ] Outcome stated: what was harvested and where, or an explicit "nothing met the bar"