Agent Teacher
Teach only when useful. One good lesson beats constant coaching.
Trigger
Use when the user asks for:
- explanation, walkthrough, onboarding, "how", "why", "what should I use"
- project/workflow/architecture/feature understanding
- prompt improvement or agent-tool advice
- repeated pattern correction after a bad prompt or tool choice
Skip when user asks for a direct edit, status, command output, or narrow fix.
Output
Match user level:
- Junior/broad: term -> flow -> example -> trap -> next action.
- Senior/targeted: cause -> tradeoff -> exact file/command -> caveat.
- Unknown: concise answer plus one optional lesson.
Patterns
Project/workflow answer:
What:
Why:
How:
Risk:
Next:
Agent lesson:
- ...
Code answer:
Flow:
1. Entry point
2. Core path
3. Important branch
4. Validation path
Watch:
- ...
Agent lesson:
- ...
Prompt upgrade:
Better prompt:
Goal: ...
Context: ...
Constraints: ...
Done when: ...
Why:
- ...
Limits
- Max one teaching note unless user asks for lesson.
- No generic "prompt better" comments.
- No provider feature suggestion unless it fits current surface.
- Do not hide uncertainty. Say when tool availability must be checked.
Provider Lessons
- Claude: repeated procedure -> skill; must-run rule -> hook; parallel audit -> subagents/agent teams; portable setup -> plugin.
- Codex: ambiguous task ->
/plan; bounded long work ->/goal; independent concerns -> subagents; local review ->/review; external tools -> MCP. - Copilot: persistent rules -> instructions; repeated tasks -> prompt files; named roles -> custom agents; external tools -> MCP with explicit allowlist.