Agentic Development

Drive work through one loop: orient, classify the task, choose the execution mode, implement, review, verify, and compound. Treat repo-local instructions and existing architecture as the source of truth. For repeated or long-running work, keep the loop supervised: fresh context each iteration, state on disk, one task at a time, hard gates, and explicit reviewer output.

Start Here

1. Run the repo scan helper if the skill bundles one; otherwise do the same discovery manually from the repo root.
2. Map the repo before touching code: instruction files, package or service boundaries, entrypoints, build and test commands, deployment surfaces, and current git state. The authoritative source of truth for project knowledge is the codebase itself and docs/ — read implemented code and documented decisions before making assumptions.
3. For frontend-heavy work, map framework boundaries, route ownership, design-system sources, shared primitives, state owners, and existing analytics or error-reporting seams before choosing an edit path.
4. Decide the task type and execution mode: implementation, review, architecture explanation, debugging, refactor, instrumentation, or release cleanup; then choose direct execution or a supervised harness loop if code will be written.
5. Choose the authoritative path for the change or explanation: the existing module, service, component, hook, page, command, model, or workflow that already owns the behavior.
6. Decide the proof before editing: what command, test, log, screenshot, trace, dependency search, or reviewer output will confirm the result.
7. Load only the reference files relevant to the current task.

Core Rules