Self-Evolving

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.

text

~/self-evolving/
├── memory.md        # HOT: stable rules, guardrails, activation cues
├── experiments.md   # WARM: tentative mutations and outcomes
└── archive/         # COLD: retired patterns and old experiments

Quick Reference#

TopicFileSetup guidesetup.mdMemory templatememory-template.mdHot memory baselinememory.mdExperiment log formatexperiments.mdEvolution cycleevolution-loop.mdSafety boundariesboundaries.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#

TrapWhy It FailsBetter MoveRewriting the whole workflow after one mistakeYou cannot isolate what actually helpedTest one mutation and compare against the previous baselinePromoting an idea after one good runLucky wins become noisy defaultsWait for three comparable wins before promotionLogging vague lessons like "be smarter"Future retrieval becomes uselessWrite the exact trigger, decision, and expected outcomeOptimizing for novelty instead of valueThe system churns without compoundingKeep only behaviors that measurably save time or reduce errorsLearning from silenceLack of complaint is not proofRequire 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 ~/self-evolving/memory.md
• Tentative experiments and outcomes in ~/self-evolving/experiments.md
• 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 ~/self-evolving/

Related Skills#

Install with clawhub install <slug> if user confirms:

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

Feedback#

• If useful: clawhub star self-evolving
• Stay updated: clawhub sync