Plan First

Long tasks fail when you start editing before you understand the code. Plan briefly, then execute.

1. Load context (before writing any code)

• Check memory: mem_context, then mem_search with the feature/module name to reuse past decisions and avoid known pitfalls.
• Read existing project architecture, rules (AGENTS.md, GEMINI.md), and similar implementations in the repository.
• Identify: entry point, data flow, test locations, and build commands.

2. Clarify only what blocks you

If a choice changes the result and cannot be inferred from the code, ask one short question. Otherwise, pick the option most consistent with the codebase and state it.

3. Write the plan (keep it short)

Goal: <one sentence>
Steps:
1. <small, verifiable change> — check: <how you will know it worked>
2. ...
Risks: <what could break>

Each step should be small enough to verify on its own. Prefer 3–7 steps.

4. Execute one step at a time

• Do step N → run its check → only then move to step N+1.
• If a check fails, fix it before continuing. Do not pile new changes on a broken state.
• If you discover the plan is wrong, stop, update the plan, and explain what changed.

5. Stay in scope

• Do not refactor, rename, or "improve" code the task did not ask for.
• Match existing style and libraries. Do not add dependencies if existing ones suffice.

6. Finish

• Run full verification (see the verify-before-done skill).
• Save architecture/design decisions with mem_save (What / Why / Where / Lesson).
• Report: what was done, what was verified, and what is left.