Harness Optimizer

Productivity

Skill that scans, scores, and auto-fixes your Claude Code harness against Anthropic's 8 harness design principles.

by CaesiumYCaesiumY/harness-optimizer

Stars4

Forks0

Watchers0

Open issues0

LicenseMIT

Last push6 months ago

Install

install

$ npx skills add CaesiumY/harness-optimizer

Update

update

$ npx skills add CaesiumY/harness-optimizer

Re-run the install command to pull the latest version of the skill.

When to use it

You have built skills, agents, commands, hooks, or a plugin and want to know if your harness follows Anthropic's long-running-app design principles.
You want a 0-100 health grade and per-principle PASS/FAIL scorecard for your .claude/ setup or plugin.json.
Your multi-agent setup lacks evaluator separation, context-reset strategy, or feedback loops and you want concrete fixes applied.
You are auditing a plugin or CLAUDE.md project before sharing it and want auto-fixes with a dry-run preview first.

Commands & flags

optimize harness: Scan, diagnose, report, and auto-fix the current project's harness.
diagnose harness --report-only: Output the PASS/FAIL/PARTIAL diagnostic report and health grade, skip auto-fix.
check my harness --dry-run: Show proposed Tier 1/Tier 2 fixes without modifying any files.
--path <path>: Target a project directory other than the current one.
harness health: Trigger phrase that returns the 0-100 health grade across the 8 principles.

How to use it

Install the skill, then trigger it with a natural-language phrase:

npx skills add CaesiumY/harness-optimizer

Then in Claude Code:

optimize harness
diagnose harness --report-only
check my harness --dry-run --path ./my-plugin

It runs four operations: Scan (auto-detects skills, agents, commands, hooks, CLAUDE.md, plugin.json, MCP config, settings), Diagnose (2-layer Grep/Glob signal collection + LLM semantic judgment against the 8 principles), Report (per-principle scores weighted to a 0-100 grade), and Fix (Tier 1 edits existing files, Tier 2 creates new files on confirmation). Use --report-only to skip fixes or --dry-run to preview them.

Why it earns its place

Codifies Anthropic's "Harness design for long-running application development" article (by Prithvi Rajasekaran) into an actionable audit, scoring your setup against 8 weighted principles like evaluator separation, context management, and feedback loops. The 2-layer diagnostic avoids naive keyword matching: Grep collects signals, then the LLM reads file contents to confirm a principle is actually implemented, cutting false positives and negatives. It is one of the few tools that audits the harness itself rather than the code, and ships tiered auto-fixes so findings translate directly into changes.

Try it

claude

> diagnose harness --report-only for this plugin and give me the per-principle scorecard