How to Build Your Own Claude Code Harness (or Buy One)

To build a Claude Code harness, you create a .claude/ folder in your project that holds your rules, skills, hooks, and settings so Claude follows your workflow every session without being re-told. Building one that you are happy with takes most developers 3 to 6 months of tuning, so the honest choice is: build it yourself if your stack is unusual or you want total control, or buy a pre-built kit (priced $0 to $149) if you want to start shipping today.

---

---

What a Claude Code harness actually is

A "harness" is just the setup that tells Claude Code how to work on your project. Without one, you re-explain your rules every chat. With one, Claude reads your instructions at the start of every session and follows them.

Anthropic describes a harness as five layers. Here is what each one is and the file it lives in:

1. Memory. Your standing rules and project facts. Lives in CLAUDE.md, a plain Markdown file Claude reads automatically at session start.
2. Tools. Extra abilities you give Claude, like a reusable skill or an MCP server (a small program that connects Claude to an outside system). Skills live in a skills/ folder.
3. Permissions. What Claude is allowed to do without asking. Lives in settings.json.
4. Hooks. Small scripts that run automatically before or after Claude does something. Hook commands are wired in settings.json and point at your own scripts.
5. Observability. A record of what happened, so you can debug. Lives in your session logs.

Get those five right and Claude behaves like a teammate who already knows your codebase.

Why "more rules" is the wrong instinct

The tempting move is to dump every rule into CLAUDE.md. That backfires. Developers report that rule-following drops to around 30% once CLAUDE.md grows past roughly 4,000 tokens (about 3,000 words). The model starts skimming. So the fix is not a longer file. The fix is better architecture: short memory, real skills for repeated tasks, and hooks for anything that must never be skipped.

Hooks are the only rule Claude cannot ignore

This is the part most people miss. A line in CLAUDE.md is a suggestion. The model can reason its way around it. A hook is law.

A PreToolUse hook runs right before Claude uses a tool. If that hook exits with code 2, the tool call is blocked. No exceptions, no negotiation. That makes hooks the correct home for quality gates ("never commit without passing tests") and guardrails ("never run a delete command on production"). If a rule truly cannot be broken, it belongs in a hook, not in your Markdown.

The third path: let Claude write its own harness

In 2026 Anthropic added dynamic workflows. For a one-off complex task, Claude can write its own short JavaScript orchestration on the fly instead of relying on your static files. It picks from six patterns: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop-until-done.

What this changes: your static .claude/ directory no longer has to encode every orchestration trick. It is now for the things that persist, like your team conventions and your stack opinions. The throwaway logic for a single hard task, Claude can generate itself.

DIY vs. Pre-built vs. Dynamic Workflow

The buy side, honestly

Three real options, with the tradeoffs stated plainly:

• Free MIT starter kit (npx). A community starter with about 27 commands, 10 hooks, and 25 skills. Costs nothing and works on any stack. You still tune it yourself.
• ClaudeKit.cc Engineer Kit, reported at $99. A paid, broad kit for general engineering work.
• Build This Now Code Kit, the $29 Code Kit. Skills and specialist agents pre-wired for a specific stack: Next.js, Supabase with row-level security on every table, and Stripe payments. It ships fast because it is opinionated. That is also the catch: opinionated kits constrain your stack choice. If you want that exact stack, you save months. If you want something else, you fight the kit.

The rule of thumb: an opinionated kit trades flexibility for speed. On a standard SaaS stack, that trade is usually worth it.

When to spend tokens on a generator-evaluator setup

For high-value work, the best architecture is two separate roles: one Claude agent generates the work, a second Claude agent grades it. This generator-evaluator split beats a single agent critiquing itself, but it costs more tokens.

Anthropic's own retro-game example shows the gap: a solo agent finished in about 20 minutes for roughly $9, while a full multi-agent harness ran about 6 hours for roughly $200. Same goal, very different bill. So save the heavy generator-evaluator setup for tasks where a better result is worth real money. For everyday work, a single agent is fine.

A decision framework

1. No .claude directory yet? Install the free starter kit today. Zero cost, instant baseline.
2. Building on a standard SaaS stack and want 48-hour shipping? A $29 to $99 opinionated kit pays for itself in the first session. You skip the skills and hooks you would otherwise spend months writing.
3. Unusual codebase, or you need control over every rule? Budget about 3 months to build and iterate your own. Keep CLAUDE.md short, put hard rules in hooks, and add skills only for tasks you repeat.
4. One-off complex task? Skip the static harness and let dynamic workflows generate the orchestration for that single job.

FAQ

How do I make Claude Code follow my rules every session?

Put your rules in a CLAUDE.md file inside a .claude/ directory at your project root. Claude reads it automatically at the start of every session. For rules that must never be bypassed, use a PreToolUse hook that exits with code 2 to block the tool call no matter what.

What is a Claude Code harness?

A Claude Code harness is the .claude/ directory structure (your CLAUDE.md rules, skills, hooks, and settings.json) that enforces your workflow, tech conventions, and quality gates every time Claude Code runs. Anthropic defines it as five layers: Memory, Tools, Permissions, Hooks, and Observability.

How long does it take to build a Claude Code harness?

Developers report that a setup they are happy with takes 3 to 6 months of iteration. A basic CLAUDE.md takes an hour, but tuning rule compliance, wiring hooks, and building skills that hold up across sessions takes sustained effort. A pre-built kit cuts this to a day or less.

Should I build my own Claude Code harness or buy one?

Build it if your stack is non-standard or you want full control. Buy one ($0 to $149) if you are on a common stack and want to ship fast, since pre-built kits include skills and hooks you would otherwise spend months writing. For one-off complex tasks, use Claude Code's dynamic workflows and let Claude generate the orchestration itself.