菜单
Evaluates agent output adversarially before committing — prevents self-assessment bias
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。
| name | adversarial-evaluator |
| description | Evaluates agent output adversarially before committing — prevents self-assessment bias |
After any agent produces output that will be committed, published, or acted on. Do NOT self-evaluate — spawn a separate evaluator instance.
The generator agent produces output in this format:
GENERATOR OUTPUT:
[the work product]
GENERATOR CLAIM:
[what the generator says was achieved]
EVALUATION CRITERIA:
[the rubric — 3-5 specific, checkable criteria]
The evaluator agent (spawned separately, no access to the generator's reasoning) must:
After the worker agent completes a task, call this skill:
Writes a structured session handoff so a fresh agent can continue without losing thread
Turn a plain-English vision or task into a verified, build-ready spec before any code. Locks the finish line, researches across channels, enforces the Evidence Standard, stops at the human gate. Adopted from CleanExpo/Fabel-Prompt-Engineer for the Authority-Site / Pi-CEO / Unite-Group build process.
Loop through advisor profiles in knowledge/board/ and return a combined multi-persona critique of a spec, draft, or decision.
Turn a plain-English vision into a verified, build-ready spec. Locks the finish line, researches across channels, enforces the Evidence Standard, and stops at the human approval gate.
Capture the user's feedback about an output and update the relevant skill or knowledge files so future outputs get sharper.
File new articles, transcripts, and notes into knowledge/ in the right place with source metadata and an evidence tag.