using-agent-skills

Using Agent Skills

Skill router for the arthurpower collection. It names the task class and the current slot in any documented sequence, then invokes that one skill. It does not do the downstream work.

User-invoked only (disable-model-invocation: true on the target): skill-authoring, self-evolving-rules. The operator must name them. Other skills, including dispatching-parallel-agents, are model-invoked playbooks the current agent may run when their triggers match.

When to use

• A new session whose first task is unclassified
• The user asks which skill / how to start / skill check / routing
• Several skills could match (ambiguous routing)
• Switching task class mid-session (planning → implementing → verifying)

Skip when the task already maps to one skill, the user already named a skill, the edit is trivial, or the work is outside this graph.

Procedure

1. Classify. Match the task to one branch in references/flowchart.md (Defining / Planning / Implementing / Verifying / Reviewing / Committing / Meta). Scale (single vs multi-session) selects a sequence, not a pile of skills in one turn. Completion: one branch label, and if scale applies, the sequence named.

2. User-invoked gate. If the match is skill-authoring or self-evolving-rules and the user did not name it, stop and ask. Completion: either a different skill, or the user named that skill.

3. Current slot. Documented sequences (SPECIFY → architecture-change-reviewer → persist → writing-plans; code-review → review-report-repair when GATE: BLOCKED) are ordered gates. Invoke this slot only; do not stack the whole sequence. Competing skills for the same slot (e.g. ACR and code-review as "the review") are not a sequence — pick one. Completion: the slot's skill name.

4. Invoke. State the routing decision in one sentence, then invoke that skill. CONTEXT / ADR writes go through domain-modeling when a persist list exists. Completion: the chosen skill started.

Promoted skill add / rename / remove / flow-placement updates this file and references/quick-reference.md in the same change. Default landing sequence: references/landing.md. Details: references/router-maintenance.md.

Registered skills (hook mirror)

hooks/pre-release-validate.ts greps this SKILL.md for each sibling directory basename. Keep every promoted skill named below. Phase / trigger tables live in references/quick-reference.md, not here.

• architecture-change-reviewer
• boundary-testing
• bounded-context-guardian
• code-review
• review-report-repair
• complexity-anti-drift
• input-contract-tests
• dispatching-parallel-agents
• domain-modeling
• error-handling-enforcer
• logicsync
• minimal-change-verifier
• self-evolving-rules
• session-handoff
• setup-arthurpower
• skill-authoring
• spec-driven-development
• systematic-debugging
• test-driven-development
• verification-before-completion
• wayfinder
• writing-plans

Acceptance

• [ ] Flowchart branch chosen before any skill invoke
• [ ] One task class; if a sequence, only the current slot invoked
• [ ] User-invoked authoring skills not auto-started
• [ ] Routing stated in one sentence
• [ ] No catch-all meta skill when a focused skill matches

References

• references/flowchart.md — branches including scale sequences
• references/quick-reference.md — phase / trigger / user-invoked vs model-invoked
• references/router-maintenance.md — add / rename / remove / flow-placement
• references/behaviors.md — optional operating habits (not routing)