error-driven-evolution
由 marsnavi 提供
Browser & Automation简介
Structured error-to-rule learning system for AI agents.
描述
该技能的功能
A structured system for converting agent errors into reusable rules stored in a lessons.md file. When a user corrects an agent or an error recurs, the agent extracts an imperative rule rather than a reflection. Rules are scanned before future decisions to prevent repeat mistakes.
Storing errors as scannable imperative rules rather than free-form notes makes them actionable at decision time, not just readable after the fact.
何时使用
- Preventing an agent from repeating the same data formatting mistake
- Capturing a correction after an agent oversteps its role boundaries
- Building a personal rule library from accumulated agent failures
- Scanning past lessons before executing a multi-step destructive operation
- Sharing anonymized mistake patterns with a community rule repository
示例工作流
User asks: The agent sent a report to the wrong audience again
- 1
Stop and identify the rule category as [COMMS]
- 2
Write an imperative rule: 'When: sending reports — Do: confirm audience scope — Don't: assume prior context carries over — Why: audience was inferred from previous session'
- 3
Append the rule to lessons.md without overwriting existing entries
- 4
Confirm briefly: 'Added to lessons: [COMMS] Confirm audience before sending reports'
- 5
Scan [COMMS] and [SCOPE] rules before next report action
Rule persisted to lessons.md; agent checks it before all future report-sending decisions
安装
clawhub install marsnavi/error-driven-evolution运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。