skills.rest
huangbaixun
huangbaixunCOMMUNITY
Explore all skills by this author →
GitHub

harness:evolve

Trim and evolve CLAUDE.md rules to fit a 60-line budget.

3|1|Updated Apr 5, 2026
Search Tags:#claude.md#token optimization#garbage collection#agent reliability#model upgrade#architecture drift#rule pruning
One-click install
npx skills add https://github.com/huangbaixun/harness-engineering --skill harness-evolve-huangbaixun
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: harness:evolve
Source: https://github.com/huangbaixun/harness-engineering/tree/main/skills/evolve
Command: npx skills add https://github.com/huangbaixun/harness-engineering --skill harness-evolve-huangbaixun

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

It helps you continuously reduce CLAUDE.md and Harness overhead as models improve, preventing documentation drift, architecture drift, and code entropy from wasting tokens and effort.

Core Features & Use Cases

  • CLAUDE.md Slimming & Rule Pruning: Evaluate each rule for whether it’s still necessary, redundant with hooks/linters, and actually tied to real failure modes—aiming to keep CLAUDE.md under 60 lines.
  • Garbage Collection for Drift: Detect documentation drift against docs/architecture.md and source code, then propose specific fixes for mismatches in rules and technology decisions.
  • Harness Streamlining After Model Upgrades: Test model behavior across a standard task set and remove or adjust scaffolding that the upgraded model now handles natively.
  • Proactive Improvement Roadmap: Upgrade from human-maintained improvements to PR-based self-healing Harness changes by logging harness-related friction and then converting it into concrete improvements.

Quick Start

Ask the AI: "Use harness:evolve to evaluate our CLAUDE.md and hooks for model-upgrade trimming, then produce a delete list with reasons and a plan to slim to 60 lines or less."

Frequently Asked Questions about harness:evolve

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I trim CLAUDE.md to reduce token waste during long iteration cycles?▼

To trim CLAUDE.md and reduce token waste, you can evaluate each rule for necessity and redundancy with hooks, mapping them to observed failure modes to keep the file under 60 lines while preserving security constraints.

What is documentation drift and how do I fix mismatches against source code?▼

Documentation drift occurs when CLAUDE.md rules and architecture decisions diverge from actual source code. You fix it by detecting mismatches against docs/architecture.md and proposing specific rule corrections to realign expectations.

How do I streamline AI agent scaffolding after a model upgrade?▼

Streamlining scaffolding after a model upgrade involves testing model behavior across a standard task set, then removing or adjusting rules that the upgraded model now handles natively to reduce maintenance burden.

When should I prune rules in CLAUDE.md for agent reliability?▼

You should prune rules in CLAUDE.md when rules become redundant with linters, are no longer tied to real failure modes, or when architecture drift causes code entropy that wastes tokens during long iteration cycles.

What is the best way to prevent architecture drift in AI harness rules?▼

Preventing architecture drift requires continuously evaluating CLAUDE.md rules against source code and docs/architecture.md, detecting technology decision mismatches, and generating a delete list with specific fixes.

Can I log harness friction and convert it into automatic CLAUDE.md improvements?▼

Yes, you can log harness-related friction during development and convert those friction logs into concrete PR-based self-healing CLAUDE.md improvements, upgrading from manual maintenance to proactive rule evolution.

Related Skills

Explore More Software Engineering Skills→
hiromailyhiromaily
126

fix-pr-review

Classify modified files and apply targeted fixes to PR review comments.

Community
Advanced
panaversitypanaversity
55

skill-validator

Validate a Skill Unit against nine quality criteria and generate a structured report.

Official
Advanced
nsalvacaonsalvacao

Automation Strategies

Automate pre-commit hooks, linters, formatters, security scans, and CI/CD pipelines across Python, JavaScript/TypeScript, Go, and universal stacks.

Community
Advanced
Srujan0798Srujan0798

self-evolve

Analyzes sprint outcomes to update CLAUDE.md, memory, and skill instructions for preventing repeat failures.

Community
Advanced