This project implements a high-level behavioral guideline and a shell-based state synchronization protocol for terminal-based AI Agents (such as Claude Code, Gemini CLI, Codex, Aider, etc.).
概览
This project implements a high-level behavioral guideline and a shell-based state synchronization protocol for terminal-based AI Agents (such as Claude Code, Gemini CLI, Codex, Aider, etc.). It prevents unverified modifications, silences fake verification theater, and forces the agent to review its own output before delivering it. redefines how AI assistants behave during software development. Instead of acting in a purely "helpful" manner — which often leads to guesswork, unchecked modifications, and unverified boilerplate — this protocol makes the agent behave like a collaborative partner: it investigates freely, asks only when the ambiguity is real, always shares its plan before touching anything, and silently reviews its own draft against a fixed checklist before delivering it. Operational state is persisted directly inside your repository so context survives terminal session restarts.
README
以下为原文 README 节选，完整内容请查看原文。
🧠 Autocritic-Skill — AI Self-Review, Check-In & State Persistence Protocol
[!NOTE] This project implements a high-level behavioral guideline and a shell-based state synchronization protocol for terminal-based AI Agents (such as Claude Code, Gemini CLI, Codex, Aider, etc.). It prevents unverified modifications, silences fake verification theater, and forces the agent to review its own output before delivering it.
💎 What is Autocritic-Skill?
Autocritic-Skill redefines how AI assistants behave during software development. Instead of acting in a purely “helpful” manner — which often leads to guesswork, unchecked modifications, and unverified boilerplate — this protocol makes the agent behave like a collaborative partner: it investigates freely, asks only when the ambiguity is real, always shares its plan before touching anything, and silently reviews its own draft against a fixed checklist before delivering it. Operational state is persisted directly inside your repository so context survives terminal session restarts.
graph TD
A["User request"] --> B{"Ambiguity level?"}
B -->|HIGH: deliverable/stack/scope unclear| C["Stop — ask max 3 questions,
or present the 4-Option Rule
if it's a design tradeoff"]
B -->|LOW/MEDIUM: minor details missing| D["Assume best practice,
investigate freely"]
C --> E["User answers"]
E --> F
D --> F["Check-in: share findings
+ planned actions"]
F --> G["User approves / adjusts"]
G --> H["Self-review: fixed checklist
+ domain lenses, silent correction"]
H --> I["Deliver polished output only"]
I --> J["Update state in
.agent/state/*.md via shell"]
🌟 Key Features
| Feature | Description | Key Benefit |
|---|---|---|
| Adaptive Ambiguity Routing | HIGH ambiguity (unclear deliverable, stack, or conflicting constraints) stops the agent to ask up to 3 questions. LOW/MEDIUM ambiguity is resolved with industry best practice, no question asked. | Eliminates both reckless guessing and pointless clarifying questions. |
| Check-In Before Acting | After investigating, the agent shares what it found and what it plans to do — and waits for your response — before modifying anything. | You catch dependencies and context the agent may have missed, before any change lands. |
| Invisible Self-Review | Before delivering, the agent re-reads its own draft against a fixed checklist and domain-specific lenses (code / text / architecture), correcting silently. The checklist is mandatory — a plain “looks fine” pass doesn’t count as a review. | Catches logical gaps, unjustified claims, and scope drift before you ever see them. |
| No Fake Testing | The agent never appends assert/console.log blocks or “if you run this…” simulations to prove its own output works — that’s theater, not verification. |
No wasted tokens on demonstrations that verify nothing. |
推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add peterson047/autocritc-skill