Skill Learning

Detect, evaluate, and suggest reusable method candidates that emerge from active work. Turns implicit expertise used during a task into explicit, shared organizational knowledge.

When to Use This Skill

• •After completing a complex multi-step task that required novel reasoning
• •During retrospectives on effort-heavy investigations or implementations
• •When the same method pattern appears across multiple tasks
• •When evaluating whether a workflow should be formalized as a skill
• •Periodically, to audit accumulated context for extraction opportunities

Methodology

Phase 1: Context Mining

Scan the current working context for repeatable method patterns. Look for:

Output: A list of 1-5 skill candidates, each described in one sentence.

Phase 2: Duplicate Check

For each candidate, verify it doesn't already exist:

1. •Compare candidate name against existing skill names (exact and synonym match)
2. •Compare candidate purpose against existing skill descriptions (semantic overlap)
3. •Check if an existing skill partially covers the candidate's scope

Decision table:

IMPORTANT: Do NOT load full skill files during this phase — compare names and descriptions only. Full skill loading wastes context budget when most candidates will be filtered out.

Phase 3: Candidate Evaluation

Score each surviving candidate on three dimensions:

3a. Effort Capitalization

How much effort was spent discovering/applying this method?

Higher effort = more value in capturing it so it's not repeated.

3b. Usage Frequency

How often would this skill be applied across future tasks?

3c. Portability

How many different agents/contexts could benefit?

Scoring Decision

Phase 4: User Presentation

Present evaluated candidates to the user for confirmation. Apply prompt-interaction-design patterns:

For 1-3 candidates: Yes/No per candidate with evaluation summary

**Skill Candidates Detected**

Based on the work done, these methods could be extracted as reusable skills:

1. **{Candidate Name}** — {one-line purpose}
   - Effort: {High|Medium|Low} | Frequency: {Recurring|Periodic|Rare} | Portability: {Universal|Multi|Single}
   - Score: {N}/9 → {Learn|Consider|Skip}
   - Create this skill? **Yes / No**

2. **{Candidate Name}** — {one-line purpose}
   ...

For 4+ candidates: Multi-select checklist

**Skill Candidates Detected**

Select which skills to create:

- [ ] **{Candidate Name}** (score {N}/9) — {one-line purpose}
- [ ] **{Candidate Name}** (score {N}/9) — {one-line purpose}
- [ ] **{Candidate Name}** (score {N}/9) — {one-line purpose}
- [ ] **{Candidate Name}** (score {N}/9) — {one-line purpose}

For candidates flagged as extensions: Suggest extending instead of creating

- **{Candidate Name}** — overlaps with existing `{skill-name}` skill
  - Recommendation: Extend `{skill-name}` rather than creating new skill
  - Extend? **Yes / No**

Anti-Patterns

• •❌ Premature extraction — Creating a skill after seeing a pattern once. Wait for the second occurrence or high effort capitalization score.
• •❌ Full skill loading during duplicate check — Wastes context budget. Names and descriptions are sufficient for Phase 2.
• •❌ Kitchen-sink skills — Bundling unrelated methods into one skill because they came from the same task. Each skill should teach one cohesive method.
• •❌ Agent-specific skills — Extracting a method only one agent would ever use. That belongs inline in the agent, not as a skill.
• •❌ Skipping user confirmation — Always present candidates for approval before creating skill files.
• •✅ Effort-aware extraction — Prioritize capturing methods that required significant effort to discover, even if usage frequency is moderate.