菜单
Author or repair an instruction layer to the Co-Aligned standard — an agent profile, agent reference, SKILL.md, skill reference, or checklist. Use when adding an agent or skill, when `coaligned instructions` flags a length breach, or when one layer has started restating another.
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
Prove whether a skill-pack change made agents better at writing code. Use when a single passing eval doesn't prove anything and you need multi-run pass@k evidence, when grading coding tasks with hidden tests the agent cannot see, or when comparing outcomes across skill-set versions.
Create design documents (WHICH/WHERE) for approved specs. A design is a max-200-line architectural sketch — components, interfaces, data flow, and key decisions with trade-offs — that gives reviewers a high-leverage point to redirect architecture before the full plan is written. Design is approved when `wiki/STATUS.md` shows the spec row at `design approved` — written there by a human signal that `kata-dispatch` or the active agent propagates.
Write implementation plans (HOW/WHEN) for approved designs. Translate an approved design into concrete steps, file changes, sequencing, and risks for a trusted agent to execute. Plan is approved when `wiki/STATUS.md` shows the spec row at `plan approved` — `staff-engineer` writes this row after a clean panel review (plans may be approved by agents).
Grade a single artifact (spec, design, plan, or implementation diff) against quality criteria and return findings by severity. Use when another skill spawns a fresh sub-agent for an independent review of its work. This skill never spawns sub-agents — it produces findings only — which structurally prevents the spec/design/plan/implement review loop from recursing.
Write specifications (WHAT/WHY) for features, changes, and improvements. Spec is approved when `wiki/STATUS.md` shows the spec row at `spec approved` — written there by a human signal (label, comment, APPROVED review, or in-session message) that `kata-dispatch` or the active agent propagates. Use when proposing changes, capturing findings as actionable specs, or evaluating spec quality. Pair with the `kata-plan` skill for the HOW side.
Merge gate for open pull requests. Verify contributor trust, classify PR type, rebase on main, fix mechanical CI failures, gate on `wiki/STATUS.md` approval state, and merge passing PRs. Sole external merge point.
| name | coaligned-layer |
| description | Author or repair an instruction layer to the Co-Aligned standard — an agent profile, agent reference, SKILL.md, skill reference, or checklist. Use when adding an agent or skill, when `coaligned instructions` flags a length breach, or when one layer has started restating another. |
The Co-Aligned architecture splits instructions across eight layers, each with
one job. A defect in one layer is a different class of problem from a defect in
another, and that separation is what makes a failed run attributable. This
skill authors and repairs the layers a contributor edits: L3–L7. The root
identity and jobs files (L1/L2) have their properties set by the standard and
are bootstrapped once by coaligned-setup —
including the CLAUDE.md section that surfaces how jobs and checklists are
discovered. Repair those there, not here.
coaligned instructions enforces a line cap and a word cap on every layer.
Either breach fails. See
references/layer-reference.md for the layers,
their properties, and their caps.
coaligned instructions flags a length breach<do_confirm_checklist goal="Verify the layer holds before committing">
coaligned instructions passes with no length findings.</do_confirm_checklist>
Name which layer you are writing — each owns one job:
Each layer's defining property and cap live in references/layer-reference.md.
If you are unsure which layer owns the content, that is the signal two layers are about to blur. Resolve it before writing.
Apply the one rule that governs every layer: no layer restates another.
When two layers must mention the same tool, separate them by voice: the lower layer describes ("ToolX sends a message"), the higher layer directs ("Use ToolX to deliver the report").
A checklist is binary verification at a natural pause point. Use the right type:
Tag every checklist so it is discoverable. See references/checklist-tagging.md for the tags and the seven properties of a good checklist.
Run coaligned instructions. On a length breach, do not just delete words —
move them to the layer that owns them:
Trimming that loses meaning means the content is in the wrong layer, not that it is too long.
coaligned instructions passes. Re-read the edited layer against its
single job: a reader following only this layer should get exactly what the layer
promises and nothing another layer owns.
coaligned instructions enforces.