Promote to Skill

Take an approach that's been validated multiple times and elevate it to a first-class Skill — queryable by intent, composable into chains, and tracked separately from one-off patterns.

When to use

• Repeated success — a pattern has reward ≥ 0.7 across 3+ episodes
• User explicit — "make this reusable", "save as a skill"
• Pre-deployment — codifying tribal knowledge before shipping the agent

Anatomy of a skill

agentdb_skill_create(
  name:         <verb-noun-noun>          // e.g. "rotate-jwt-refresh-token"
  description:  <one-liner>               // what it does
  precondition: <plain English query>     // matched by intent embedding
  action:       <step list>               // executable steps
  outcome:      <success criteria>        // what "worked" looks like
  tags:         [<topic>, <stack>, ...]
  metadata:     { sourceEpisodeIds: [...], confidence }
)

The precondition is what makes the skill discoverable — it's embedded and matched against future task intents. Write it like a search query: "need to rotate a JWT refresh token without invalidating active sessions", not like a docstring.

Composition

Skills can chain via agentdb_skill_compose:

compose(
  intent: "set up authenticated API with token rotation",
  available: [skill_a, skill_b, skill_c]
) → bandit picks the best chain ordering

The bandit tracks composition rewards over time; bad orderings decay automatically.

Don't

• Don't skill-ify single successes. One data point doesn't generalize.
• Don't skill-ify project-specific code paths. Skills should be transferable; if the action only works in one repo, it's a pattern, not a skill.
• Don't put secrets or repo-specific paths in action. Use placeholders.