revfactory Claude Code Harness — Summary

This is a research repository, not a user-facing framework: it contains a controlled A/B experiment comparing Claude Code output quality with vs without a pre-configured .claude/ harness across 15 software engineering tasks at three difficulty levels. The central finding, reported in a bundled paper, is that harness pre-configuration improves average quality scores from 49.5 to 79.3 (+60%), with the effect scaling with task complexity (Expert +36.2 points). The actual harness delivered is small — 4 slash commands (/experiment, /evaluate, /report, /run-advanced-experiment), 3 skills (experiment-runner, output-evaluator, report-generator), and an experiments/ directory with YAML test cases, worktree-isolated baseline and harness agents, and JSON result tracking. There is no installable plugin or reusable end-user tool — the value is the research evidence. Differs from seeds: this is the only "methodology proof" in the catalog rather than a tool — analogous to agent-os (markdown scaffold + proof of concept) but focused on empirical validation rather than opinionated guidelines. The experiment design resembles the parallel-fan-out orchestration of superpowers and revfactory-harness but instrumentalized for measurement.

revfactory Claude Code Harness — Overview

Origin

GitHub: https://github.com/revfactory/claude-code-harness
Stars: 98
Language: HTML (paper visualizations)
Last commit: 2026-03-06
No license declared.

Philosophy

This repository treats "harness" not as an installable tool but as an empirical claim: that providing structured pre-configuration to Claude Code demonstrably improves output quality. The README presents itself as a research paper with methodology, results tables, and dimension-by-dimension analysis.

From the README:

"In a controlled A/B experiment across 15 software engineering tasks, Harness improved average quality from 49.5 to 79.3 points (out of 100) — a 60% improvement. The effect scales with complexity: Basic +23.8, Advanced +29.6, Expert +36.2."

"Harness lives in the .claude/ directory and provides four types of guidance: architectural blueprint (CLAUDE.md), skills (domain-specific knowledge), agents (role decomposition), and commands (workflow orchestration)."

Key Design Decisions

1. Research-first: the repository exists to prove a claim, not to ship a product.
2. Parallel agent teams: the experiment uses BaselineAgent and HarnessAgent running in parallel on the same task, then EvaluatorAgent comparing outputs — a three-agent experiment harness.
3. 10-dimension scoring: quality measured on completeness, code quality, efficiency, accuracy, and structure (each 0-10, total 100).
4. Worktree isolation: baseline and harness experiments run in separate worktrees to prevent contamination.

revfactory Claude Code Harness — Architecture

Distribution

Not installable as a user-facing plugin. This is a research repository. The .claude/ configuration functions as a self-contained experiment environment.

Directory Tree

claude-code-harness/
├── .claude/
│   ├── CLAUDE.md              # Project memory / experiment context
│   ├── commands/
│   │   ├── experiment.md      # /experiment — run A/B test with agent teams
│   │   ├── evaluate.md        # /evaluate — compare outputs
│   │   ├── report.md          # /report — generate summary report
│   │   └── run-advanced-experiment.md
│   └── skills/
│       ├── experiment-runner.md
│       ├── output-evaluator.md
│       └── report-generator.md
├── experiments/
│   ├── cases/                 # YAML test case definitions
│   ├── results/               # Per-case baseline/ and harness/ directories
│   └── reports/               # Aggregated reports
├── paper/
│   └── figures/               # Charts used in the paper
└── README.md                  # Acts as the research paper

Required Runtime

• Claude Code (no version specified)
• No additional dependencies

Target AI Tools

Claude Code only (experiment design is Claude Code-specific).

revfactory Claude Code Harness — Components

Commands (4)

Skills (3)

Agents (3, experiment-specific)

These are NOT defined as .claude/agents/*.md files — they are instantiated inline within the experiment command:

Test Cases (15 documented)

revfactory Claude Code Harness — Uniqueness

differs_from_seeds

No seed is purely a research harness — this is the only framework in the catalog whose primary deliverable is empirical evidence rather than an installable tool. The closest analogy is agent-os (Archetype 4: markdown scaffold, zero primitives) but agent-os ships reusable markdown templates, while revfactory-cc-harness ships a measurement apparatus. The experiment design borrows the parallel-fan-out pattern from what superpowers and revfactory-harness implement, but instrumentalized for controlled comparison rather than production use. It is most notable as the only framework that provides quantitative proof that .claude/ pre-configuration improves LLM coding quality.

Positioning

This is a proof-of-concept / research paper repository. The "harness" is both the experimental apparatus and the subject of the experiment. Not suitable for production use as a standalone tool — the value is the validation data.

Observable Failure Modes

1. No install path: there is no way to use this as a standalone harness outside the experimental context.
2. Korean-only documentation: README and skill files are in Korean; most users cannot evaluate the quality without translation.
3. Inline agent prompts: agents are defined as inline strings within command files, not as reusable .claude/agents/*.md definitions — they cannot be reused outside the experiment context.
4. No license: no license declared, making derivative use legally uncertain.
5. Dormant: last commit 2026-03-06; likely not actively maintained.