Enhance Skills

enhance-skills

Turn corrections into durable skill improvements. Pick the mode from the request:

• Harvest (default) — "enhance the skills I used", "record what you learned",

"don't make that mistake again". Scan this conversation and write lessons.

• Retrofit — "add self-learning to ``", or automatically when Harvest

meets a skill whose SKILL.md has no ## Self-learning block.

Lessons store

Every skill's lessons resolve to one of two places; resolve per skill and use the same resolution for reading and writing:

• In-place — directory writable (source checkout or --plugin-dir copy):

sibling LESSONS.md; committed here, lessons ship to everyone.

• External — read-only or update-clobbered install (under

.../plugins/cache/): $HOME/.agent-data/ctx42-skills/lessons//.md, $HOME-rooted to survive plugin updates.

Resolve the directory from the base dir the host reports. Read both files when both exist and obey the union; write only to the writable one (create parent dirs). Never write under .../plugins/cache/ even if the OS says it's writable — a plugin update wipes it.

When the resolved store is a writable sibling LESSONS.md and the external file also exists, offer to merge the external lessons into LESSONS.md (dedup per Lesson format) and delete the external file — lessons collected on a read-only install otherwise stay stranded there.

Harvest

1. List skills used. Collect every skill actually invoked in this

conversation (a /: call or Skill tool use). Ignore skills only mentioned, never run.

1. Mine each for lessons. For each used skill, find where the user corrected

the agent, or the agent caught its own mistake, while that skill was driving. A lesson is a concrete rule that would prevent a recurrence — not a one-off fact about this task. Skip praise, restated requirements, and task-specific detail.

1. Resolve stores. For each used skill, resolve its SKILL.md path and its

[lessons store](#lessons-store). If the path is unknown or ambiguous, ask.

1. Propose. Present, per skill, the resolved target file and the one-line

rules you intend to add. Wait for confirmation before writing.

1. Write. On approval, for each skill: retrofit the ## Self-learning block

if the SKILL.md is in-place and lacks it, then append each new rule to the resolved store (see [Lesson format](#lesson-format)), deduplicating against lines already there.

1. Report. Per skill: the file written and each rule added. Nothing else.

Retrofit

Install the ## Self-learning block as the last ## section of the target SKILL.md — append it at the end of the file, so the job owns the top of the body and the constant block sits last. Skip if the block already exists. Fill ` and from the target's path. Do not create any lessons file; the store appears lazily on the first lesson. Retrofit edits SKILL.md`, so it needs an in-place checkout — on a read-only install, report that and stop.

Block to insert (verbatim, with / resolved to the real names):

## Self-learning

Read this skill's lessons and obey them: sibling LESSONS.md, else $HOME/.agent-data/ctx42-skills/lessons//.md when this directory is read-only. On a correction or self-caught mistake, append a one-line rule to whichever is writable (creating it) and report where.

Lesson format

A lessons file is a flat list of imperative rules, one per bullet, newest last:

# Lessons

Rules learned for the `` skill. Read before running; obey each line.

• One rule per bullet, imperative, terse — the bar of a style rule.
• Conform to the repo markdown style: wrap at ~80 columns, continuation lines

indented two spaces.

• Concrete and general: "Derive the commit type from the diff, not the branch

name" — not "the user wanted fix not feat here".

• No dates, ticket ids, or task-specific nouns. A lesson outlives its task.
• Dedup: fold a near-duplicate into the existing rule; never add a twin.
• Create the file with the # Lessons header on the first lesson; append after.

Output

Report tersely: no preamble or narration; state each fact once; don't restate output the user can already see. Per touched skill, list the file written and the rules added (or the block installed). Always name the resolved path so the user knows whether a lesson shipped (sibling LESSONS.md) or stayed local (.agent-data); never fabricate a write you could not perform.

Self-learning

Read this skill's lessons and obey them: sibling LESSONS.md, else $HOME/.agent-data/ctx42-skills/lessons/craft/enhance-skills.md when this directory is read-only. On a correction or self-caught mistake, append a one-line rule to whichever is writable (creating it) and report where.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ctx42
• Source: ctx42/skills
• License: MIT

Install and usage instructions live in the source repository linked above.