Harness — The Team-Architecture Factory
Turns a one-sentence domain description into a coordinated agent team and the skills they use, automatically building an evolvable automation system for Claude Code projects.
The skill writes only project-local files (.claude/agents, .claude/skills, CLAUDE.md); no credential theft, covert exfiltration or malware; Phase 0 mandates an audit report and user confirmation of the plan, and _workspace artifacts are preserved for auditability. Deductions: no least-privilege constraints on generated agents (full tool access by default), mandatory model:"opus" imposes high cost, no rollback/recovery for CLAUDE.md modifications, confirmation only implicit at initial build, sensitive-data handling unaddressed.
SKILL.md and six references are internally consistent, with templates, error-handling tables, test scenarios and clear branching (new/extend/maintain) plus a Phase selection matrix. Deductions: static review only — no committed test suite or CI proving key-path reproduction; output quality depends heavily on model behavior; concrete failure-feedback examples for abnormal input are thin.
Triggers are explicit in the description (build/extend/audit/sync), non-goals (.claude/commands) and mode-selection boundaries are stated, and user-expertise adaptation is a plus. Deductions: description and body are Korean only — no Chinese/English trigger support; core function depends entirely on Claude Code experimental agent teams and Anthropic services, with mainland-China reachability risk; very high token cost (opus mandated for all agents) is not declared as a boundary.
Apache-2.0 license is explicit; README provides version (1.2.0), multilingual install, structure and FAQ; the skill is well layered (SKILL.md + six references with ToCs and progressive-disclosure rules). Deductions: no CHANGELOG or in-skill versioning, no declared maintenance ownership or update path, publisher unverified; cross-file numeric consistency (e.g. team-size tables) unverified.
README presents author-measured A/B data (+60% quality, 15/15 win rate, n=15) honestly labeled as author-measured with third-party replication pending, plus many copyable usage prompts. Deductions: all effectiveness evidence is author-claimed; static review cannot verify generated outputs are directly usable; the +60% experiment is not independently reproducible and cost/benefit against very high token spend is unproven.
Auditable primary material exists: full SKILL.md, six references, concrete bug cases (7 SatangSlide boundary bugs) and methodology; fact vs inference largely separated and the FAQ discloses evidence limits. Deductions: no third-party replication, no committed test results or CI artifacts, the A/B experiment lives in a sister repo not statically checkable here; evidence types are single-source (author claims). Static cap of 5 applies; scored 4.
- This is a static source-only review; nothing was executed, and reliability/effectiveness scores are capped accordingly.
- Generated agents get full tool access by default and opus is mandatory, so token costs can be high — review generated agent definitions and set a budget before use.
- The skill modifies project CLAUDE.md and creates many files; run under version control to enable rollback.
- Core function depends on Claude Code experimental agent teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) and Anthropic services; mainland-China reachability is uncertain.
- Description and body are Korean only, which may limit triggering and comprehension for non-Korean users.
- The +60% quality claim is author-measured (n=15) with no third-party replication; base adoption decisions on your own pilot.
- Publisher is unverified by FollowSkills; identity is unknown.
What does this skill do, and when should you use it?
Harness is a Claude Code plugin (and meta-skill) that converts your domain description into a team of specialized agents plus the skills they use. It picks from six pre-defined team-architecture patterns — Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, and Hierarchical Delegation — and generates agent definitions under .claude/agents/, skills under .claude/skills/, and an orchestrator skill that ties the workflow together. The system is designed to evolve: after each run it collects feedback and keeps agents, skills, and the CLAUDE.md change log continuously updated.
On trigger, it first audits the project's existing .claude/agents/, .claude/skills/, and CLAUDE.md, detecting drift and classifying the run as new build, extension, or maintenance. It then analyzes the domain, selects an execution mode (agent teams by default; subagents or hybrid as alternatives) and an architecture pattern, and generates agent definition files — always file-based, with role, principles, I/O protocols, and team communication protocols, always using model: "opus". It generates one or more skills per agent using Progressive Disclosure (SKILL.md under 500 lines, details moved to references/), plus an orchestrator skill that specifies data-passing protocols (message/task/file/return-value), error handling, and _workspace/ intermediate artifact management. It registers a minimal pointer and change history in CLAUDE.md, and finishes with structural validation, trigger verification (including should-NOT-trigger near-miss queries), and with-skill vs without-skill comparison tests.
- Researchers who need an agent team to investigate a topic from multiple angles, cross-validate findings, and produce a comprehensive report
- Full-stack developers who want a coordinated pipeline for design, frontend (React/Next.js), backend API, and QA testing from wireframe to deployment
- Content creators building teams for webtoon episode production or YouTube planning (writing, character design, panel layout, SEO)
- Code review leads who want parallel agents checking architecture, security, performance, and style, merged into one report
- Teams with an existing harness performing maintenance, agent/skill synchronization, extension, or architecture changes
- Technical writers generating API documentation from a codebase, with usage examples and completeness review
What are this skill's strengths and limitations?
- Six mature architecture patterns cover sequential, parallel, generation-review, and hierarchical collaboration shapes
- Mandatory file-based agent definitions and a CLAUDE.md change log ensure cross-session reuse and traceability
- Built-in QA methodology: with-skill vs without-skill comparison, trigger verification with near-miss queries, incremental QA
- An author-run A/B experiment (n=15) reports average quality rising from 49.5 to 79.3, with honest disclosure of its limits
- Well documented: trilingual README plus references/ with templates and five real team configurations
- Deeply tied to Claude Code-exclusive mechanisms (TeamCreate, SendMessage, TaskCreate, Agent tool, CLAUDE.md); not directly portable to other runtimes
- Mandates model: "opus" for every agent call, implying high API cost with no documented lighter option
- The +60% quality figure comes from a 15-task author-measured A/B; no third-party replication exists
- Requires the experimental CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 flag, which may change
- No standalone automated test suite in the repo; verification relies on the skill's runtime validation workflow
How do you install this skill?
Option 1 (Marketplace): run /plugin marketplace add revfactory/harness, then /plugin install harness@harness-marketplace in Claude Code. Option 2 (Global skill): run cp -r skills/harness ~/.claude/skills/harness. Prerequisite: enable the experimental Agent Teams feature with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.
How do you use this skill?
Trigger with natural language in a Claude Code session, e.g. "Build a harness for this project" or "Design an agent team for this domain" (Korean "하네스 구성해줘" and Japanese triggers are also listed). The skill runs a seven-phase workflow: status audit → domain analysis → team architecture design → agent definition generation → skill generation → integration & orchestration → validation & testing, then enters a continuous evolution loop with feedback capture and change logging.
How does this skill compare with similar options?
The README explicitly positions it against neighbors: coleam00/Archon is a Runtime-Configuration Factory on the same L3 Meta-Factory layer (choose Archon for deterministic runtime configs); SaehwanPark/meta-harness is a Codex port of the same concept; affaan-m/ECC is a standardization layer across harnesses (different layer, serial combination possible); wshobson/agents is a subagent/skill catalog whose entries can be absorbed as parts inside a Harness-generated team.