Process
Phase 1 — Clarify
- Use
grill-with-docsskill to get context. - Ask clarifying questions about scope, behavior, and approach.
- Skip obvious questions.
- If invoked from automation (e.g.
saml-todo-next): skip interactive questions, use provided context directly.
Phase 2 — Write Plan
- Create plan file via
saml-planfileskill. - Write a self-contained plan detailed enough for a cheap model to execute.
- Use
tddskill to plan tests. - Plan must include:
- File paths and function names to create/modify.
- TDD: which tests first, what each validates.
- Key decisions and rationale.
- Dependencies/prerequisites.
- Error handling and edge cases.
- State assumptions explicitly. If uncertain, ask.
- Simplicity first. No overengineering.
Phase 3 — Review-Fix Loop
Launch a general purpose sub-agent (fresh context window) to review the plan. Pick the model using saml-model skill.
The reviewer must NOT share context with the planning phase — pass all context explicitly in the prompt:
- The full plan file contents
- Relevant codebase context (AGENTS.md, project conventions)
- The original user request/goal
You are a senior engineer. You HATE this plan and want to criticize it. Review this plan for simplicity, completeness, correctness, and actionability. Return PASS or FAIL with specific issues. Write review to
PLANFILE.review.md.
- If PASS → done.
- If FAIL → fix the plan file, re-run review.
- Max 10 iterations. If still failing, report to user with path to
PLANFILE.review.md.
Done
Tell user: "✅ Plan written: <path>"
If iteration limit hit, show remaining issues and point to PLANFILE.review.md.