Claude Harness Generator

v1.0.0 | Built for Claude Opus 4.8 | MIT

Generate a production-ready, best-practice Claude Code harness -- fully tailored to your use case, in minutes, at any skill level. Or validate, update, and improve a harness you already have.

A harness is a complete Claude Code setup for a project: CLAUDE.md + rules + specialized agents + slash-command skills + memory + scoped permissions + self-learning. Normally you hand-craft all of that over days and still end up with contradictions, bloat, and broken references. The Harness Generator interviews you (or takes a preset), then builds the whole thing -- coherent, validated, and immediately usable.

Quick Start

# 1. Download this project
git clone https://github.com/Popschlock/Claude-Harness-Generator.git
cd Claude-Harness-Generator

# 2. Launch Claude Code in the directory
claude

# 3. Create your harness -- then just answer the prompts
/create

That is the whole flow. Answer a short interview -- your domain, tools, team shape, how autonomous Claude should be -- and the Harness Generator writes a complete, tailored harness to a path you choose. No Claude Code expertise required: the interview does the work, and everything generated follows current best practices.

It tailors to whoever you are:

• Solo dev or enterprise team -- memory tier, roles, and coordination scale to team size
• With or without source control -- Git, Perforce, or none
• With or without MCP servers / external tools -- only the integrations your intake justifies, never speculative
• One focused project or a multi-area hub -- several related work areas sharing a parent configuration layer
• Any domain -- software, data, DevOps, game dev, research, legal, finance, hiring, support, and more

The full run takes about 10-20 minutes including the interview.

What it does

1. Create a new harness. Pick a preset (one of 4 base profiles or 16 ready-to-use bundled domains -- fast) or request a custom build where the architect synthesizes a reusable profile for your exact use case. The Generator then designs the architecture, generates every file in five focused passes, and validates the result before handing it to you.

2. Improve a harness you already have.

Command	What it does
/validate-environment	Structural + quality check of an existing harness (references resolve, frontmatter valid, sizes, routing, hub registry)
/upgrade-environment	Audits an existing harness against current best practices, interviews you about pain points, and implements the improvements you approve (including single <-> multi-area-hub conversions)
/update	Refreshes the Generator's own best-practices knowledge base (web research, or a local-only mode that just ingests the ProvideKnowledge folder)

Prerequisites

• Claude Code CLI
• An Anthropic API key or a Claude subscription

Presets

The intake offers a preset path (fast) or a custom path (more tokens/time, best for long-lived harnesses) where the architect synthesizes a reusable DOMAIN_PROFILE.md. Docs/StarterProfiles/DOMAIN_REFERENCE.md maps ~20 common domains to a recommended starting point.

Base profiles (slim, ~110-170 lines each)

Profile	For
Software Development	Web apps, APIs, libraries, CLI tools
Knowledge Work	Legal, research, finance, writing
Data & Analysis	Accountants, analysts, reporting, dashboards
DevOps & Infrastructure	DevOps engineers, SRE teams, platform engineers

Bundled domain profiles (16)

Ready-to-use, adversarially audited domain profiles in Docs/DomainLibrary/: api-design, book-publishing, course-design, customer-support, data-engineering, data-science, financial-modeling, game-development, grant-writing, hiring-pipeline, legal-research, llm-app, market-research, product-management, security-audit, social-media. The high-stakes domains (legal, finance, security, support, hiring) ship with the relevant professional/regulatory safeguards built in.

Don't fit any preset? The custom path builds a configuration from scratch via a deep interview.

What gets generated

your-project/
|-- CLAUDE.md                    # AI assistant behavior and routing
|-- .claude/
|   |-- settings.json            # Permissions and tool access
|   |-- rules/                   # Orchestration, autonomy, context, self-learning, errors
|   |-- agents/                  # Specialized AI agents for your domain
|   |-- skills/                  # Commands: /state-save, /state-load, /update, /health-check
|-- Docs/
    |-- GETTING_STARTED.md       # Plain-language guide for your team
    |-- Memory/                  # Project knowledge (scales with your team)
    |-- State/                   # Session snapshots for save/resume
    |-- Retro/                   # Self-learning observations and proposals
    |-- Environment/             # Generation metadata and evolution log

Generated harnesses are immediately effective out of the box. The built-in self-learning system (/update) then refines routing, agents, and rules based on your actual usage, closing in on a hand-tuned configuration within weeks.

Key generated commands

Command	What it does
/state-save	Saves session progress so you can close and resume later
/state-load	Restores context from a saved session
/update	Analyzes usage patterns and proposes improvements
/health-check	Validates harness integrity and reports issues

Architecture

The Harness Generator uses an orchestrator + subagent pattern with 5 agents:

• Orchestrator (CLAUDE.md + rules): routes requests, coordinates the pipeline, presents results
• Intake Interviewer (Sonnet): deep project interviews when presets don't fit
• Environment Architect (Opus): designs the architecture from intake answers + best practices; picks a team-architecture pattern and execution mode; writes DOMAIN_PROFILE.md for custom builds
• Component Generator (Opus): writes harness files in five focused passes
• Environment Validator (Sonnet): runs the quality checklist + skill-triggering tests, a Phase-0 drift audit, and boundary-crossing checks
• Upgrade Analyzer (Opus): compares an existing harness against best practices and writes tiered recommendations

All agent output is written to disk (artifact-first), keeping the main context lean.

Project structure

Claude-Harness-Generator/
|-- CLAUDE.md                           # Generator behavior
|-- .claude/
|   |-- rules/                          # 4 rules (creator-core, intake-protocol, generation-standards, quality-gates)
|   |-- agents/                         # 5 agents (interviewer, architect, generator, validator, upgrade-analyzer)
|   |-- skills/                         # 4 skills (create, validate-environment, upgrade-environment, update)
|-- Docs/
    |-- AgentGuidelines/                # Best-practices knowledge base (18 topics)
    |-- AgentPlaybooks/                 # 6 step-by-step process guides (+INDEX)
    |-- Templates/                      # Annotated reference implementations (52 files)
    |-- StarterProfiles/                # 4 slim base profiles (+ PROFILE_FORMAT, DOMAIN_REFERENCE)
    |-- DomainLibrary/                  # 16 bundled domain profiles
    |-- ProvideKnowledge/               # Drop zone for user-contributed knowledge

How it works (technical)

1. The /create skill verifies the target directory and writes CREATION_CONTEXT.md
2. The orchestrator runs intake -- experience level, work-area shape, preset-vs-custom (or a deep interview)
3. Intake answers are written to GENESIS.md in the target directory (multi-area hubs also get HUB_GENESIS.md)
4. The environment-architect designs the full architecture (ARCHITECTURE.md); custom builds also write a reusable DOMAIN_PROFILE.md
5. The component-generator writes files in five passes (foundation -> agents -> skills -> infrastructure -> docs); hubs add a shell pass
6. The environment-validator runs the quality checklist (see Docs/Templates/References/validation-guide.md)
7. The orchestrator presents the result with smoke-test instructions

Contributing

See CONTRIBUTING.md for how to add starter profiles or bundled domains, create or update templates, improve the knowledge base, and test changes.

License

MIT. See LICENSE.