Configuration-as-Code for Claude Code

Make Claude Code
Truly Yours

Harness transforms Claude Code from a stateless tool into an intelligent, context-aware collaborator that understands your project, respects your conventions, and improves over time.

Get Started See How It Works
7
Configuration Pillars
13
Atomic Skills
21
Maturity Score
96%
Eval Score (w/ skills)

The Harness Difference

Experience the leap from raw automation to engineered precision.

Without Harness

  • Stateless prompts require constant context re-provisioning.
  • Inconsistent naming conventions and architecture patterns.
  • No guardrails against destructive or unsafe actions.
  • No feedback loop for long-term intelligence growth.
  • Fragmented developer experience across environments.

With Harness

  • Automatic context injection from deep project scanning.
  • Enforced architectural pillars that Claude strictly follows.
  • Atomic skills library for complex, multi-step tasks.
  • Continuous learning through evaluation and feedback logs.
  • Unified configuration-as-code shared across the team.
The Framework

7 Pillars of Configuration

Harness organizes every aspect of Claude Code configuration into seven foundational pillars. Each is independently auditable and improvable.

01 // COMPOSITION
bolt

Skill Composition

Define discrete atomic capabilities. Chain complex behaviors through modular skill blocks that can be versioned and tested independently.

What should Claude be able to do?
02 // DATA
database

Context Engineering

Precision knowledge injection. Manage how data is chunked, prioritized, and served to the model within token constraints.

What does Claude need to know?
03 // FLOW
account_tree

Orchestration & Routing

Dynamic intent detection and multi-agent delegation. Route queries to the most efficient model or specialized skill.

How should work be distributed?
04 // MEMORY
memory

Persistence & State

Long-term memory management. Store user preferences, past interactions, and operational state across sessions.

What should Claude remember?
05 // TRUST
verified_user

Quality Gates

Automated evaluation layers. Run validation scripts, schema checks, and filters on every output before it ships.

How do we catch mistakes early?
06 // POLICY
gavel

Permissions & Safety

RBAC for AI agents. Define what actions are permitted and which require human-in-the-loop authorization.

What can Claude do without asking?
07 // UX
psychology

Ergonomics & Trust

Tone management and interaction tuning. Control how the model expresses uncertainty and presents responses.

How should Claude communicate?
The Toolkit

13 Atomic Skills

Each skill is a focused, composable unit that implements one or more pillars. Call them individually or let the router dispatch automatically.

hub
harness-engineer Core chevron_right
Master router — dispatches to all 7 pillar skills based on intent

Master router that classifies intent and dispatches to the right specialist skill. Handles simple lookups directly and coordinates multi-pillar work by sequencing atomic skills. The single entry point for all harness engineering.

Routes to all skills
harness-engineer scaffold context memory permissions hooks routing ergonomics
rocket_launch
harness-init Core chevron_right
Bootstrap orchestrator — 5-phase setup from scaffold to context

Bootstraps a complete harness by orchestrating a five-phase sequence. Analyzes the project, generates CLAUDE.md, configures permissions, initializes memory, installs hooks, and validates the context budget — all in one pass.

Bootstrap sequence
harness-init 1. scaffold → 2. permissions → 3. memory → 4. hooks → 5. context
query_stats
harness-audit Core chevron_right
7-pillar evaluator — scores 0-3 per pillar, produces improvement report

Evaluates an existing harness against all 7 pillars, scoring each from 0 (absent) to 3 (advanced). Produces a structured report with the total maturity score, identifies the weakest pillar, and recommends prioritized improvements with direct skill routing.

Standalone — evaluates existing state
autorenew
harness-loop Core chevron_right
Continuous improvement — audit, fix weakest pillar, validate, repeat

Continuous improvement engine that cycles: audit the harness, identify the weakest pillar, route to the atomic skill that fixes it, validate the improvement didn't regress others, log what changed, and repeat.

Improvement cycle
harness-loop audit improve any atomic skill validate
construction
harness-scaffold Atomic chevron_right
Generates tailored CLAUDE.md and .claude/ directory from project analysis

Analyzes codebase tech stack, git history, engineering patterns, and architecture to generate a project-tailored CLAUDE.md and .claude/ directory structure. The foundational artifact that all other skills build upon.

→ Context Engineering
data_usage
harness-context Atomic chevron_right
Optimizes token budget and restructures CLAUDE.md for signal density

Optimizes what information Claude loads and when. Analyzes token budget, eliminates wasted context, restructures CLAUDE.md for progressive disclosure, and ensures the most critical instructions are front-loaded.

→ Context Engineering
psychology
harness-memory Atomic chevron_right
Designs cross-session memory with typed files, index, and staleness auditing

Designs and maintains the cross-session memory system — MEMORY.md index, typed memory files (user, feedback, project, reference), staleness auditing, and proper organization so future conversations have complete context.

→ Persistence & State
admin_panel_settings
harness-permissions Atomic chevron_right
Configures tool permission boundaries using blast radius principle

Configures permission boundaries in settings.json using the blast radius principle: auto-allow reads and local builds, prompt for shared-state mutations. Detects tooling and tunes specificity to eliminate unnecessary prompts.

→ Permissions & Safety
webhook
harness-hooks Atomic chevron_right
Auto-format, lint, secret scan, command gating, test-on-change hooks

Designs and implements automated hooks that fire before/after tool use or at conversation end. Provides patterns for auto-formatting, linting, secret scanning, command gating, test-on-change, and system health checks.

→ Quality Gates
widgets
harness-skills Atomic chevron_right
Analyzes workflows and decomposes into composable skill systems

Analyzes project workflows and decomposes them into a well-composed skill system. Discovers workflow candidates, classifies by complexity, designs atomic and composed skills, checks for overlaps and gaps, and recommends priorities.

→ Skill Composition
device_hub
harness-routing Atomic chevron_right
Designs subagent delegation, parallelization, and worktree strategies

Designs agent orchestration strategies including subagent delegation, parallelization patterns, build-loop workflows, worktree isolation, and mega-skill routing. Produces concrete orchestration plans tied to project constraints.

→ Orchestration & Routing
verified_user
harness-gates Atomic chevron_right
Quality gates at post-edit, pre-commit, pre-push, and conversation-end

Designs quality gates and feedback loops at multiple stages: post-edit, pre-commit, pre-push, pre-merge, and conversation-end. Implements self-evaluation patterns, regression detection, and automated feedback loop closure.

→ Quality Gates
tune
harness-ergonomics Atomic chevron_right
Tunes verbosity, autonomy, trust level, and communication style

Tunes human-AI interaction by configuring trust levels, output style, status reporting, decision surfacing, error communication, and personality. Captures feedback memories to improve interaction patterns across sessions.

→ Ergonomics & Trust
Workflows

How It Works

Three primary workflows cover the full lifecycle — from initial setup to ongoing optimization.

1

Scaffold

harness-scaffold

Analyze project structure, frameworks, and patterns. Generate a tailored CLAUDE.md and settings.json that aligns with your architectural requirements.

2

Configure Permissions

harness-permissions

Set up tool-specific permission boundaries. Auto-allow safe operations, prompt for destructive ones.

3

Set Up Memory

harness-memory

Initialize the persistent memory system. Seed with user profile, project context, and key decisions.

4

Install Hooks

harness-hooks

Add automated quality gates: code formatting, linting, secret scanning, and test triggers.

5

Optimize Context

harness-context

Fine-tune what Claude loads. Trim waste, front-load critical info, reference files instead of inlining.

CLAUDE.md 
# Build & Test
pnpm dev   # start dev server
pnpm test  # execute vitest

# Architecture
- Next.js 14 + App Router
- PostgreSQL via Prisma

# Conventions
- Use "server actions" sparingly
- PascalCase for Components

# Critical Rules
1. NEVER update prisma/migrations
2. ALWAYS run tests first

Ready to bootstrap?

Start your project with a fully configured harness in seconds.

Get Started
1

Read All Harness Files

Scan CLAUDE.md, settings.json, memory system, hooks, and skill definitions.

2

Score Each Pillar

Rate each of the 7 pillars from 0 (not configured) to 3 (advanced). Total maturity out of 21.

3

Produce Report

harness-audit

Generate a structured report with pillar scores, top 3 recommendations, and direct routing to the skill that fixes each gap.

CLAUDE.md 
# Build & Test
pnpm dev   # start dev server
pnpm test  # execute vitest

# Architecture
- Next.js 14 + App Router
- PostgreSQL via Prisma

# Conventions
- Use "server actions" sparingly
- PascalCase for Components

# Critical Rules
1. NEVER update prisma/migrations
2. ALWAYS run tests first
1

Audit Current State

Run a full 7-pillar audit to identify the current maturity score.

2

Identify Weakest Pillar

Find the pillar with the lowest score. This is where improvement has the highest impact.

3

Improve It

Route to the atomic skill for that pillar. Make targeted, validated improvements.

4

Validate & Repeat

harness-loop

Re-score to confirm improvement. Check for regressions. Continue to the next weakest pillar.

CLAUDE.md 
# Build & Test
pnpm dev   # start dev server
pnpm test  # execute vitest

# Architecture
- Next.js 14 + App Router
- PostgreSQL via Prisma

# Conventions
- Use "server actions" sparingly
- PascalCase for Components

# Critical Rules
1. NEVER update prisma/migrations
2. ALWAYS run tests first
Measurement

Maturity Scoring

Every harness is evaluated across seven core dimensions. Scores range from 0 (not configured) to 3 (advanced), providing a clear roadmap for improvement.

Skill Composition 3 / 3
Context Engineering 3 / 3
Orchestration 2 / 3
Persistence 3 / 3
Quality Gates 2 / 3
Permissions 3 / 3
Ergonomics 2 / 3
Example audit output — total score: 18/21
Quick Start

Get Started

Begin your journey with these foundational steps.

01

01 Install Skills

Copy the skills/ directory into your project.

02

02 Run the Engineer

Call harness-engineer to start the analysis.

03

03 Bootstrap

Run harness-init for fresh projects to set defaults.

04

04 Audit

Run harness-audit for existing setups to check compliance.

05

05 Iterate

Use harness-loop for continuous improvement. Establish a recurring cycle of refinement to maintain top-tier maturity scores across all pillars.

Launch Feedback Loop