compass

Files (1)

• SKILL.md

SKILL.md (first lines)

# Compass

`compass` is a top-level user-invocable supplementary orientation skill. Use it for bounded strategic orientation before implementation, not implementation itself.

It reads and defers to `reference/routing-guidance.md` before final route recommendations. It hands implementation to the selected route or domain skill.

## When to use

- The request has an unclear goal, broad scope, or several plausible routes.
- The team needs a short option comparison before picking a category, local skill, helper, or planning step.
- The next move may be a built-in planner, `review` skill, `architecture-integration`, a domain skill, or a narrower implementation route, but the fit is not clear yet.
- The best answer is a decision brief, not code, edits, tests, or rollout work.

## When not to use

- The route is already clear and the selected domain skill can start.
- The work is a direct implementation, validation, review, research, git, or UI critique task.
- A full plan is needed. Use the built-in `plan` agent or the `implement` skill.
- A hard quality or architecture challenge is needed. Use `critique` or `architecture-integration`.
- Implementation is already complete and needs a finish pass. Use the `review` skill or `impeccable`.
- The request belongs directly to `architecture-integration` or a domain skill without ambiguity.

## Orientation workflow

1. Reframe the user's goal in one or two plain sentences.
2. List constraints, unknowns, risks, and decisions that affect the route.
3. Read `reference/routing-guidance.md` before final route recommendations, keeping the skill families and refinement layer boundaries intact.
4. Name 2-3 viable directions that could reasonably satisfy the goal.
5. Compare those directions by value, cost, reversibility, risk, and evidence gaps.
6. Recommend one direction and the next route, skill, helper, or planning step.
7. Hand implementation to the selected route or domain skill.

## Output shape

- Goal reframe: restate the strategic target and the user-visible outcome.
- Constraints, unknowns, risks, decisions: separate what is fixed from what still needs judgment.
- Viable directions: give 2-3 viable directions, each with a short label and fit summary.
- Comparison: compare each direction by value, cost, reversibility, risk, and evidence gaps.
- Recommended direction: choose one path and explain why it beats the alternatives.
- Next step: name the next route, skill, helper, or planning step.
- Success criteria: define what must be true when the chosen path is done.
- Verification criteria: define how the chosen path should be checked.
- Question: ask at most one material question only when the answer changes the route or plan.

## Bounded reflection rules

- Keep the pass compact enough to help the next actor move.
- Prefer a clear recommendation when the available evidence is enough.
- Treat unresolved context as an evidence gap, not a reason to expand scope.
- Do not implement, edit files, run broad verification, or create plans from this skill.
- Do not invent new routes, helper IDs, support tiers, or validated surfaces.

## Collaboration in this repo

- Use the built-in `plan` agent or `implement` skill when the next step needs a concrete multi-step plan.
- Use research tools or the built-in Explore behavior (or equivalent research behavior) or `Librarian` / Context7 when local patterns or source-of-truth context are missing.
- Use `critique` or `architecture-integration` when the recommendation needs a harder challenge pass.
- Use domain skills for implementation once the route is chosen.
- Use the `review` skill after substantial implementation, not as a substitute for orientation.

## Guardrails

- `compass` is not a primary route, not a routing-guidance replacement, and not an implementation executor.
- It is not a replacement for the built-in planner, `review`, `implement`, `critique`, `architecture-integration`, or any domain skill.
- It keeps route advice separate from public support or validation claims.
- It recommends one next move rather than creating a parallel backlog.
- It stops after orientation and hands work to the selected owner.

From HanTechnology/grok-toolkit, MIT.

More from HanTechnology/grok-toolkit