Using Skills

1. Choose the workflow

Choose the applicable branch using the host's native skill selection:

• If one skill clearly matches, load its full instructions and follow them before acting. Do not use this router for a clear single-skill task, load every skill, or force a full lifecycle.
• If work spans workflows or no single primary skill clearly applies, choose the primary procedure from the routing table, then load only supporting procedures that materially apply. An explicit user skill selection takes precedence as the primary workflow.
• If the collection or a selected skill is unavailable, say so and use a scoped fallback. Do not claim to have used a skill based only on its name, description, or plan link.

Follow user requests and repository instructions. Skill links route work; they do not install dependencies, authorize actions, or spawn agents. Read anti-ai-writing before drafting human-facing text.

Keep each skill's boundary: planning does not implement; review does not repair; preference prediction does not establish correctness; CI configuration does not establish release readiness; a draft does not authorize publication. The calling workflow resumes already-authorized work after a supporting skill finishes.

For an accepted Plan-to-Build request, use implement-plan as the coordinator; writing-clean-simple-code and domain skills support it rather than replacing its verification and final independent review gates. Jira intake supplies source intent, agent-planning supplies tasks, and implement-plan verifies execution against both. The generated plan must contain the execution handoff, not just a chat reminder, so native Build and separate-session execution can carry the coordinator instruction. When authoring a plan, identify only the skills that materially apply, link each one from the plan's location, map it to the task or decision it supports, and state the specific guidance the executor should use. Keep the implement-plan coordinator instruction distinct from task-specific skills. Tell the executor to read each linked skill before its task; do not assume host discovery or skill aliases will activate it. If a linked skill is unavailable, surface that limit and use a scoped fallback without claiming the skill was applied.

When skills reference each other, reuse already-loaded guidance and evidence rather than recursively restarting workflows. In testing/debugging, retain the current reproduction and task scope. Do not re-run unchanged verified work solely because a handoff occurred; reopen coverage after relevant changes or new contradictory evidence.

If a required skill or tool is unavailable, state the limitation and use a scoped fallback when possible. Never claim a missing procedure, independent review, live check, or deployment occurred. Use delegation only when permitted and useful, with raw intent and the exact candidate; do not create personas merely to route tasks.