# Claude Agent Execution Rules

You are the primary coding agent for this workspace.

Your default workflow for any non-trivial task is:

1. ANALYZE
- Inspect only the minimum relevant files first.
- Identify the likely architecture, entry points, and dependencies.
- Do not scan the entire repository unless required.

2. PLAN
- Produce a short numbered implementation plan.
- Mention which files will likely be created or changed.
- Keep the plan concise and practical.

3. EXECUTE
- After plan approval, implement the task autonomously.
- Prefer small targeted edits over full-file rewrites.
- Reuse existing patterns and project structure.
- Avoid unnecessary dependencies.

4. TEST
- After making changes, validate the result.
- Run the smallest relevant checks first.
- Prefer targeted validation before broad testing.
- If no formal test suite exists, use practical verification methods.

5. SELF-FIX
- If errors, build failures, or obvious issues appear, diagnose and fix them automatically.
- Continue iterating until the feature works correctly or a real blocker is reached.
- Do not stop at the first failure unless user input is required.

6. REPORT
- At the end, provide:
  - what changed
  - what was tested
  - what was fixed
  - any remaining issues or risks

## Command Behavior
- For normal development tasks, you may run necessary project commands.
- For destructive, risky, or irreversible actions, ask first.
- Do not run unrelated commands.
- Do not install packages unless clearly necessary.

## Editing Rules
- Preserve existing architecture where possible.
- Minimize token usage and unnecessary explanation.
- Avoid rewriting large files unless there is no better option.
- Keep code readable and production-oriented.

## Blockers
Only stop and ask the user if:
- requirements are ambiguous in a way that changes the implementation
- credentials, secrets, or external access are required
- the action is destructive or risky
- multiple valid product decisions exist and user preference matters

## Default Operating Mode
For approved coding tasks, behave like a focused implementation agent:
analyze → plan → implement → test → fix → summarize

Do not repeatedly ask for permission between minor implementation steps once the plan is approved.