By name: skill-auto-discovery effort: low description: "After completing a complex task, check whether the sequence of steps used represents a recurring pattern worth codifying into a skill. Logs skill candidates to the bus for weekly review by Dane." triggers: ["skill candidate", "repeated pattern", "should this be a skill", "codify this", "skill discovery"]
Skill Auto-Discovery
After any task that required 5+ tool calls or a non-obvious sequence of decisions, run this checklist. The goal is to catch repeating workflows before they calcify into tribal knowledge.
When to Run
Run automatically after completing any task where:
- Total tool calls exceeded 5
- You made a non-trivial routing decision not covered by existing skills
- You handled an edge case by improvising (vs. following a defined rule)
- You found yourself repeating a sequence you've done before
Also runs as part of the weekly reflection scan (see self-evaluation-triage skill).
Detection Checklist
Before logging a skill candidate, verify it meets at least 2 of these 4 criteria:
| # | Criterion | Check |
|---|---|---|
| 1 | Same sequence of tool calls used 2+ times in the last 30 days | grep daily memory files for similar task patterns |
| 2 | Decision tree with 3+ branches applied manually more than once | was there a multi-step if/then that didn't map to any existing skill? |
| 3 | Vendor selection heuristic applied 3+ times without a written rule | did you pick a vendor based on unwritten knowledge? |
| 4 | Edge case handled outside any existing skill | something new happened and you figured it out on the fly |
If 0 or 1 criteria match: log nothing. Not every task needs a skill.
Logging a Skill Candidate
When 2+ criteria match:
cortextos bus log-event quality skill_candidate info \
--meta "{
\"proposed_name\": \"<kebab-case-name>\",
\"description\": \"<one sentence: what it does>\",
\"trigger_tasks\": [\"<task_id_1>\", \"<task_id_2>\"],
\"evidence\": \"<what pattern was observed>\",
\"effort_estimate\": \"low|medium|high\",
\"agent\": \"blue\"
}"
Then write a one-line entry to today's memory file:
- SKILL CANDIDATE: <proposed_name> — <description>. Evidence: <tasks>. Effort: <low/medium/high>
Do NOT build the skill yourself. Dane reviews candidates weekly and decides what to build.
Weekly Surface (runs as part of weekly reflection scan)
On Monday morning, pull all skill_candidate events from the past 7 days:
cortextos bus list-events --type quality --subtype skill_candidate --since 7d --format json
If any candidates exist, send a message to Dane:
cortextos bus send-message dane normal "Weekly skill candidates from Blue: <count> candidates logged this week. Top candidates: <list proposed_name + description for top 3>. Full list in activity log."
If no candidates: no message needed.
What Makes a Good Skill
A skill candidate is worth building when:
- It would save >10 min of reasoning per occurrence
- It occurs at least 2–3x per week
- It has clear inputs and outputs
- It does not require David's judgment to execute
A candidate is NOT worth building when:
- It's a one-off edge case
- It requires judgment that belongs to David
- An existing skill already covers it with minor extension
Guardrails
- Never propose a skill that bypasses an approval workflow
- Never propose a skill that touches tenant personal data without explicit rules
- Candidates are suggestions only — Dane decides whether to build
- Do not log the same pattern twice in one week
Pair with self-evaluation-triage for post-meld reflection. Pair with the tasks skill for tracking what patterns produce the most backlog.