Capture
Detect when a recent moment is reusable knowledge, recommend the best place to capture it, ask for approval, and then update the approved destination safely.
Trigger
Activate only when at least one of these is true:
- The user invokes this skill directly
- The conversation includes an explicit capture cue such as:
- "remember this"
- "this should be a skill"
- "we always forget this"
- "the correct command is"
- A workflow reaches a checkpoint such as:
- a fix is confirmed
- a decision is finalized
- a non-obvious command or parameter is established
Do not activate continuously or speculatively. This skill is a checkpoint helper, not ambient monitoring.
Instructions
- Confirm the trigger moment and identify the candidate knowledge from recent context.
- Read
references/target-selection.mdto decide whether the best destination is:- an existing skill
- a new skill
tasks/lessons.mddocs/solutions/...- ignore
- Before recommending a destination, check for an existing capture to avoid duplication.
- Search matching skills with
rg - Check
tasks/lessons.md - Check
docs/solutions/if it exists
- Search matching skills with
- Recommend one destination with a short reason and confidence statement.
- If confidence is high for an existing skill, name that skill directly.
- If the decision is ambiguous, present the top two options instead of bluffing certainty.
- Ask for approval using plain numbered choices so the flow works across platforms.
- Example:
- Approve recommended destination
- Choose a different destination
- Ignore for now
- Example:
- If the user approves a write, read
references/edit-policy.mdand follow the target-specific rules.- Existing skill: apply append-only updates directly after approval, but preview structural changes first.
- New skill: create a minimal package under
skills/<skill-name>/plus a rootREADME.mdentry. - Lessons: append one short rule to
tasks/lessons.md. - Solution docs: create
docs/solutions/on demand and write one compact solution document.
- Keep the update focused.
- Prefer one primary destination per moment.
- Add cross-links only when they materially help.
- Do not write to multiple destinations unless the user explicitly asks.
- Summarize what changed and reference the edited files.
Guardrails
- Do not write anything before approval.
- Do not force a skill update when a lesson is the cleaner fit.
- Do not rewrite a skill's core workflow without showing a preview first.
- Do not create a new skill for a one-off project fact.
- Do not claim certainty when the destination is genuinely unclear.
References
references/target-selection.md- destination heuristics, confidence rules, and duplicate checksreferences/edit-policy.md- mixed policy for skill edits plus write templates for all destinationsreferences/examples.md- sample trigger moments, approvals, and destination outcomes