github.com
feat(skills): add prose-self-check skill — 4-layer AI-vernacular audit before agents emit prose artifacts
目标用户 · developer
阅读原文 ↗feat(skills): add prose-self-check skill — 4-layer AI-vernacular audit before agents emit prose artifacts: ## Context Agents in this repo emit long prose artifacts (session-log narratives, ADR context sections, retrospectives, PR descriptions). The repo enforces **code** quality (CodeQL, complexity, coverage), **structural**
个性化解读(发生了什么 / 为什么重要 / 影响 / 建议)将在接入 LLM 后按你的角色生成。
痛点信号
- this issue ports a proven 4-layer ai-vernacular self-check model (running in production in a sibling hermes agent) into ai-agents as a reusable, agent-invocable skill.
- this is a capability addition, not a bug-fix; there is no claim that current agent output is measurably sloppy (a grep of `.agents/` artifacts for slop tells found ~none).
- **semantic.** emptiness gate: name the disagreeable claim each paragraph makes; if you can't, it's fluent filler ("saying nothing at length" — a top-5 tell no scanner sees).
- higher reliability (present every turn), but duplicated across agents and harder to version.
- ## out of scope - distributional **classifiers** (pangram/gptzero) — cannot run in-agent; only the burstiness proxy is in scope.
#developer#agent#issues