Match Existing

Step 2: Compare semantically

For each existing skill, compare:

• Are the triggers similar? (same commands, same file patterns, same context)
• Are the errors prevented similar? (same root cause, same symptoms)
• Are the actions similar? (same fix, same workaround)

Step 3: Determine verdict

DUPLICATE (similarity > 0.95)

This experience is already covered by: [skill-name]

  Existing skill: [skill description]
  Your experience: [extracted pattern]

  Nothing new to add.

  [OK, cancel]  [No, it's different — force new]

AMENDMENT (similarity 0.7 - 0.95)

This looks like it improves: [skill-name] (v[version])

  EXISTING SKILL:
  ✓ [existing action 1]
  ✓ [existing action 2]
  ✓ [existing error prevented]

  YOUR EXPERIENCE ADDS:
  + [new action or detail]
  + [new error prevented]
  + [new trigger condition]

  [Complete this skill]  [It's a new skill]  [Cancel]

NEW SKILL (similarity < 0.7)

No matching skill found. This will be contributed as a new skill.

  Name suggestion: [auto-generated-name]
  Category: [auto-detected domain]

  [Continue]  [Cancel]

Step 4: Hand off

• If DUPLICATE + user accepts: stop, no further action
• If DUPLICATE + user forces: treat as NEW, proceed to anonymize-session
• If AMENDMENT + user confirms "Complete": proceed to anonymize-session with amendment context (target skill + additions)
• If AMENDMENT + user says "New skill": proceed to anonymize-session as new skill
• If NEW: proceed to anonymize-session

Output

Pass to anonymize-session:

• verdict: "new" | "amendment"
• target_skill: skill name (only if amendment)
• target_version: skill version (only if amendment)
• extracted_pattern: the full extracted experience
• additions: what's new vs existing (only if amendment)

Restrictions

• Do NOT skip this step: even if the user thinks it's new, always check
• Do NOT compare with pipeline skills: only compare with d9 experience skills
• The developer always has the final say: if they say it's different, treat it as new
• Show the full diff for amendments: the dev must see exactly what they're adding

Self-Check

• All existing skills in skills/ were read and compared?
• Pipeline/meta skills were excluded from comparison?
• The verdict was clearly communicated to the user with actionable options?
• For amendments, the diff between existing and new is clearly shown?
• The developer made an explicit choice before proceeding?