Self-Evolving自我进化

When to Use

User wants the agent to improve a repeated workflow without blind self-rewrites. The skill handles local experiment logs, promotion of proven patterns, and explicit value gates before a new behavior becomes stable.

Architecture

Memory lives in ~/self-evolving/. If ~/self-evolving/ does not exist, run setup.md. See memory-template.md, memory.md, experiments.md, evolution-loop.md, and boundaries.md for the operating model.

CODEBLOCK0

Quick Reference

Topic	File
Setup guide	INLINECODE8
Memory template

memory-template.md | | Hot memory baseline | memory.md | | Experiment log format | experiments.md | | Evolution cycle | evolution-loop.md | | Safety boundaries | boundaries.md |

Requirements

• - No credentials required
• No extra binaries required
• No network access required

Core Rules

1. Start From Real Friction

• - Evolve only after a failed attempt, repeated correction, or measurable bottleneck.
• Do not invent mutations just because a task feels interesting.

2. Change One Lever at a Time

• - Test one prompt pattern, decision rule, retrieval step, or file habit per experiment.
• Small mutations make the winning variable obvious.

3. Gate by Value, Not Novelty

• - Promote a pattern only when it improves speed, quality, or reliability across at least three comparable uses.
• Unproven ideas stay tentative in experiments.md.

4. Keep Local Evidence

• - Record the trigger, mutation, outcome, and next action for every experiment.
• Tell the user before the first persistent write that this skill keeps concise local notes for repeat improvement.
• Promote durable rules into memory.md only after evidence repeats.

5. Prefer Promotion Over Rewrite

• - Convert winners into short rules, checklists, or retrieval triggers.
• Stable systems compound by accumulation, not by starting over.

6. Respect Hard Boundaries

• - Follow boundaries.md before storing data or changing behavior.
• Never modify the installed skill files, exfiltrate unrelated data, or run hidden experiments on the user.

Common Traps

Trap	Why It Fails	Better Move
Rewriting the whole workflow after one mistake	You cannot isolate what actually helped	Test one mutation and compare against the previous baseline
Promoting an idea after one good run

Lucky wins become noisy defaults | Wait for three comparable wins before promotion | | Logging vague lessons like "be smarter" | Future retrieval becomes useless | Write the exact trigger, decision, and expected outcome | | Optimizing for novelty instead of value | The system churns without compounding | Keep only behaviors that measurably save time or reduce errors | | Learning from silence | Lack of complaint is not proof | Require explicit feedback or repeated success evidence |

Security & Privacy

Data that leaves your machine:

• - None by default

Data that stays local:

• - Stable rules, guardrails, and activation notes in INLINECODE17
• Tentative experiments and outcomes in INLINECODE18
• First-time local storage should be announced before the first write

This skill does NOT:

• - Call external APIs
• Read or store credentials
• Modify its own installed instructions
• Read unrelated files outside the active task plus INLINECODE19

Related Skills

Install with clawhub install <slug> if user confirms:

• - self-improving — learn from corrections and compound execution quality over time
• INLINECODE22 — keep durable long-term context and retrieval patterns
• INLINECODE23 — compare options and commit to a clear next move
• INLINECODE24 — structure deliberate practice and feedback loops
• INLINECODE25 — follow through on next steps once a better pattern is chosen

Feedback

• - If useful: INLINECODE26
• Stay updated: INLINECODE27