菜单
Use when training setup, checkpoint evaluation, ablation, metric computation, or other new empirical output is required. Not for ordinary production-code changes that create no new experiment or evaluation result.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | autopilot-lab |
| description | Use when training setup, checkpoint evaluation, ablation, metric computation, or other new empirical output is required. Not for ordinary production-code changes that create no new experiment or evaluation result. |
| argument-hint | <task description> [--mode setup|eval|auto] [--parent <slug>] [--ref <similar-model-path>] [--intensity direct|quick|standard|strong|thorough|adversarial] [--report] [--from spec|scaffold|run|eval|summary] |
| metadata | {"group":"entry","fam":"code","invocation_class":"entry-router","modes":["setup","eval"],"blurb":"Rapid experiment prototyping around training setup and checkpoint evaluation/analysis.","use_when":"Use when training setup, checkpoint evaluation, ablation, metric computation, or other new empirical output is required.","not_for":"Not for ordinary production-code changes that create no new experiment or evaluation result."} |
This is a compact pre-approval entry router. Its manifest-owned frontmatter is the authoritative discovery metadata; it intentionally contains no execution procedure.
Use only this router, harness-manifest.json, and core/WORKFLOW.md §0.2 to
propose a route. Present the one-time confirmation in core/WORKFLOW.md §0.4
unless the same route and scope are already approved. Do not load references
before approval.
After approval, direct/quick sessions and the standard+ dispatch-depth-1 owner load
capabilities/autopilot-lab.md, then use the Reference Index below. Assigned
stage workers load only their assigned stage contracts.
| File | Load when | Obligation |
|---|---|---|
references/owner-execution.md | After approval, by the selected direct/quick session or dispatch-depth-1 owner | Read the complete execution procedure before material work. This is the router's only post-approval reference edge. |
Follow the portable artifact, worktree, role, and verification guards in the selected owner contract. Runtime projections must report unsupported mechanics and must not claim physical instruction masking, token, billing, or cost savings without verified evidence.
Use when source code must be implemented, debugged, refactored, or code-audited through a plan, execution, test, and report loop. Not for requirements-only changes, new empirical experiments, or document-only editing.
Use when persistent analysis of code, a paper, or a document must be created or refreshed because it is absent, stale, or explicitly requested. Not for read-only project orientation, context recovery, or status reporting.
Use when durable cross-project user preferences must be inferred from coding, writing, or analysis evidence and stored as a profile. Not for one-project context recovery, casual preference acknowledgment, or ordinary task execution.
Use when a visual product surface needs references, design tokens, components or mockups, review, and development handoff. Not for implementing an already-approved design in code or for document prose work.
Use when a new paper, presentation, report, proposal, or other user-facing document must be produced from evidence. Not for correcting only an existing document or for source-code implementation.
Use only when autopilot-draft dispatches document strategy and evidence-plan creation. Not for top-level user requests or primary capability routing.
