# Agent Instructions

This repo is mostly Markdown process artifacts plus small standard-library Python scripts.

Preserve concise, actionable wording.

Do not add unnecessary framework complexity.

Do not copy from external skill repos.

Keep each skill self-contained.

When editing a skill, update related templates if needed.

## Route-first context loading

Resolve the task route before loading workflow guidance.

- Load only the selected skills and directly relevant current-state or project files by default.
- Do not preload the full skill pack merely because it is installed.
- Expand context when the task spans routes, a compatibility seam requires it, or verification exposes missing context.
- Searches and reads remain allowed unless explicitly forbidden. This rule limits the default working set, not discovery.
- Treat current project state as more authoritative than templates and examples.

See `docs/context-isolation.md` for the full loading and expansion rules.

Run validation after changes:

```bash
python -m py_compile scripts/validate_repo.py
python -m py_compile scripts/install_claude_code.py
python scripts/validate_repo.py
```

Prefer small diffs.

Preserve the MIT license and acknowledgments.

Do not add generated marketing fluff.

When editing handoff behavior, preserve `Workflow state`: active modes, current phase, current loop, next gate, context risk, and active hypothesis. Do not expand handoff into a transcript summary.

## Quality standards

Every skill must include YAML frontmatter:

- `name`
- `description`

Every skill must include:

- Purpose
- When to use
- Inputs
- Workflow
- Outputs
- Stop conditions
- Anti-patterns

Every template must be short and directly usable.

Every example must show a concrete use case.

Every install path must remain portable across tools.

Claude Code support is first-class.

Codex support is maintained.

Manual install must remain possible because skills are plain folders with `SKILL.md` files.