| name | skill-picker |
| description | Choose the most appropriate repo-local skill for an ambiguous task by consulting .agents/skills/README.md. |
| category | context-docs |
| kind | atomic |
| phase | context |
| requires_write | false |
| requires_slurm | false |
| requires_benchmark_artifacts | false |
| delegates_to | [] |
| output_schema | skill_run_summary.v1 |
Skill Picker
When to use
Use this skill when the user asks which repo-local skill applies, gives a broad or ambiguous
workflow request, or asks for routing across issue, PR, benchmark, SLURM, artifact, or docs work.
Purpose
Select the smallest useful repo-local skill stack when the user request is ambiguous.
Workflow
- Read
.agents/skills/README.md and .agents/skills/skills.yaml.
- Classify request type with the routing matrix below.
- Choose one primary skill; add secondary skills only when phases are distinct.
- Prefer registry metadata over prose when routing conflicts:
category, kind, phase, requires_write, requires_slurm,
requires_benchmark_artifacts, and delegates_to.
- If a specific user choice conflicts and is clearly wrong, override with rationale.
- Report selected skills and continue execution if requested.
Routing matrix
| User intent | Primary skill | Secondary skill |
|---|
| "Take next issue" | goal-issue-implementation | gh-issue-autopilot |
| One selected issue to PR | gh-issue-autopilot | implementation-verification, gh-pr-opener |
| Ambiguous or underspecified issue | issue-contract-maintainer | gh-issue-sequencer |
| Create new issue | gh-issue-creator | gh-issue-sequencer |
| Sequence issue batch | gh-issue-sequencer | gh-issue-priority-assessor |
| Fix PR comments | gh-pr-comment-fixer | pr-ready-check |
| Open PR | gh-pr-opener | artifact-provenance |
| Verify implementation | implementation-verification | pr-ready-check |
| Branch cleanup | clean-up | pr-ready-check |
| Worktree setup or cleanup | skill-picker | gh-issue-autopilot, clean-up; see AGENTS.md |
| Benchmark-sensitive review | review-benchmark-change | benchmark-row-status |
| Camera-ready benchmark audit | analyze-camera-ready-benchmark | artifact-provenance |
| Submit training or benchmark job | slurm-campaign-submit | artifact-provenance |
| Submit issue-791 Auxme job | auxme-issue791-submit | slurm-campaign-submit |
| Stage external data | data-staging-provenance | artifact-provenance |
| Synthesize multiple evidence sources | evidence-synthesis | paper-facing-docs |
| Durable context note | context-note-maintainer | paper-facing-docs when claims matter |
Negative routing
- Do not use
autoresearch for ordinary cleanup.
- Do not use
paper-facing-docs for non-claim docs.
- Do not use
gh-issue-autopilot for ambiguous issues; route to issue-contract-maintainer.
- Do not use
auxme-issue791-submit for non-issue-791 campaigns.
- Do not use
benchmark-overview as validation proof; it is an orientation skill.
- Do not combine multiple skills that own the same phase unless one is a legacy compatibility entry.
Guardrails
- Do not treat skill-picker as mandatory pre-step for explicit single-skill requests.
- Avoid proposing overly broad bundles that overlap the same phase.
- Use exact repository-specific terminology from AGENTS/skills docs.
- Match mutating skills to the user's permission and the current worktree safety constraints. For
worktree creation or cleanup, route to the existing issue/cleanup skill and the
AGENTS.md
worktree lifecycle sections instead of inventing a new skill.
Output
Compact routing note:
- Selected skill(s)
- Why they fit
- Skipped alternatives
- Next action