revfactory Harness — Summary

Harness by revfactory is a "meta-factory" Claude Code plugin: a single skill (harness/SKILL.md) that, when invoked, analyzes the user's domain and generates a custom agent team with specialized skill files tailored to that domain. Rather than being a harness itself, it is a harness generator — the output is .claude/agents/ and .claude/skills/ files configured for the user's specific project. The framework defines 6 architectural team patterns (Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, Hierarchical Delegation) and guides the selection based on domain analysis. Agent teams are the default; all agents use model: "opus" for maximum quality. The framework self-identifies as "L3 Meta-Factory / Team-Architecture Factory" and explicitly references ECC (L2 cross-harness) and Archon (L3 runtime-config factory) as co-existing tools. Stars: 3,552. Differs from seeds: no seed is a harness generator — all seeds ARE harnesses. The closest analogy is BMAD-METHOD (persona-based multi-agent) but BMAD ships fixed personas, while this generates custom ones per domain from scratch.

revfactory Harness — Overview

Origin

GitHub: https://github.com/revfactory/harness
Stars: 3,552
License: Apache-2.0
Language: HTML (documentation)
Last commit: 2026-05-15
Multilingual: English, Korean, Japanese

Philosophy

From README:

"Harness is a team-architecture factory for Claude Code. Say 'build a harness for this project' and the plugin turns your domain description into an agent team and the skills they use — picked from six pre-defined team-architecture patterns."

The key insight: "harness" is not a set of pre-built skills/commands but an architecture that generates them. The framework treats agent team design as a learnable craft with named patterns, not a free-form task.

Self-Positioning

From README:

"Harness lives at the L3 Meta-Factory layer of the Claude Code ecosystem — the layer that generates other harnesses rather than being one."

Layer taxonomy explicitly called out:

• L3 Meta-Factory / Team-Architecture Factory (Harness itself)
• L3 Meta-Factory / Runtime-Configuration Factory (Archon)
• L2 Cross-Harness Workflow (ECC/affaan-m)

Manifesto-Style Quotes

"Inside L3, we pick a specific sub-layer: Team-Architecture Factory." "All agents use model: 'opus'. Harness quality is directly tied to agent reasoning ability." "The harness is not a static artifact — it is an evolving system. Incorporate feedback after every run."

Key Design Decisions

1. Agent teams as default: 2+ agents → team-first; subagents only when team overhead exceeds benefit.
2. Opus everywhere: all generated agents default to the most capable model.
3. Iterative evolution: the harness is registered in CLAUDE.md for cross-session continuity; updated after each run.
4. Phase 0 audit: before generating anything, audit existing harness state and detect drift.

revfactory Harness — Architecture

Distribution

Claude Code plugin marketplace:

/plugin marketplace add revfactory/harness
/plugin install harness@harness

Also installable as a direct global skill:

cp -r skills/harness ~/.claude/skills/harness

Directory Tree

harness/
├── .claude-plugin/
│   └── plugin.json             # Plugin manifest
├── .github/
├── skills/
│   └── harness/
│       ├── SKILL.md            # Main 6-phase workflow
│       └── references/
│           ├── agent-design-patterns.md   # 6 architectural patterns + decision trees
│           ├── orchestrator-template.md   # Team/subagent orchestrator templates
│           ├── team-examples.md           # 5 real-world team configurations
│           ├── skill-writing-guide.md     # Skill authoring guide
│           ├── skill-testing-guide.md     # Testing + evaluation methodology
│           └── qa-agent-guide.md          # QA agent integration guide
├── _workspace/                 # Generated workspace artifacts
├── docs/
├── README.md
└── README_KO.md / README_JA.md

Required Runtime

• Claude Code (plugin marketplace)

Target AI Tools

Claude Code (primary). The meta-factory generates agents/skills for Claude Code's native formats.

revfactory Harness — Components

Skills (1 + 6 references)

Generated Output (per invocation)

6 Architectural Patterns

Execution Modes (generated agents)

revfactory Harness — Uniqueness

differs_from_seeds

No seed is a harness generator — all 11 seeds ARE harnesses. The closest is BMAD-METHOD (Archetype 1, persona-based multi-agent with named subagents) but BMAD ships fixed personas while this generates custom domain-specific ones. Also resembles superpowers in using skills-only behavioral framework (single skill, no commands), but superpowers' skills enforce development discipline while this skill generates other skills. The self-aware L3/L2/L1 layer taxonomy explicitly positions this relative to ECC (named as a co-existing L2 tool), making it unique as the only framework in the catalog that explicitly maps its own position in a framework ecosystem.

Most Distinctive Features

1. Meta-factory: generates other harnesses rather than being one.
2. Opus mandate: all generated agents hardcoded to the most capable model.
3. 6-pattern taxonomy: explicit names for team architecture patterns used as a design vocabulary.
4. Drift detection: Phase 0 audit prevents harness rot over time.

Observable Failure Modes

1. Cost: all-opus generated teams are expensive — each agent spawned will cost more than a haiku or sonnet invocation.
2. Korean-first documentation: the primary SKILL.md is in Korean; reference materials may be Korean-only.
3. Generated harness quality depends on generation quality: a poorly described domain produces a poor harness.
4. CLAUDE.md coupling: the harness pointer in CLAUDE.md can conflict with other plugins that also modify CLAUDE.md.