Claude Code Boilerplate vs the Code Kit: Harness, Not Template

A Claude Code boilerplate is a starter codebase, usually Next.js with auth and payments, plus a CLAUDE.md file that gives Claude basic project context. A harness is different: it is the layer of skills, agents, workflows, and a lean CLAUDE.md that tells Claude how to build every feature, so the work keeps moving without you steering each step. Anthropic named this distinction in its June 1, 2026 post "A harness for every task," and its research shows the harness shifts output quality by 22 points while changing the underlying model moves it by only 1.

---

---

Boilerplate vs harness: the one idea to take away

A boilerplate is scaffolding you write code against. You still prompt Claude for every feature, and you still check its work. A harness is the instruction layer that sits between you and Claude. It carries the rules, the specialist agents, and the build steps, so Claude knows how to ship a feature the same way every time.

Why this matters to you: if you start from a boilerplate, you are the orchestrator. You decide the order, you catch the mistakes, you repeat yourself. If you start from a harness, Claude does that orchestration. You describe what you want and review the result.

Anthropic's own numbers make the case. In its "A harness for every task" post, the harness architecture accounted for a 22 point swing in output quality on its internal benchmark, while swapping the model behind it moved the score by 1 point. The instructions around the model matter more than the model.

Why a long CLAUDE.md hurts you

CLAUDE.md is the file Claude Code reads at the start of every session to learn your project's rules. Most people think a longer file means a smarter assistant. The opposite is true.

Anthropic recommends keeping CLAUDE.md under 200 lines. Experienced harness builders aim for under 50. Here is the plain reason: Claude reads the whole file before you type a single word, and every line spends part of its limited reading budget (its "context window," the amount of text it can hold at once, about 200,000 tokens).

The "IFScale" instruction-following benchmark found that models start dropping rules once you stack 150 to 200 of them into one prompt. So a 1,000 line AGENTS.md, the kind some WordPress and SaaS boilerplates ship, does two bad things at once. It burns context before the first question, and it gives Claude more rules than it can reliably follow. The $29 Code Kit ships a pre-tuned CLAUDE.md under 50 lines on purpose. The depth lives in skills that load only when a task needs them, not in one giant file.

Dynamic workflows make a hand-written agents.md counterproductive

Claude Code shipped dynamic workflows on May 28, 2026 in version 2.1.154. Here is what they do in plain terms: instead of you writing a fixed script of "first do this agent, then that agent," Claude writes its own JavaScript orchestration script on the fly. That script runs in the background, separate from your chat, and can spawn up to 16 agents in parallel and up to 1,000 agents per run. You trigger it with the keyword "ultracode" or by asking Claude to "use a workflow."

This changes the math. A static boilerplate's hand-written agents.md tries to hard-code an order of operations. But Claude now plans that order itself when it has a good harness around it. So a long, fixed instruction file is not just heavy, it fights against the feature that makes Claude fast. Give Claude the right tools and rules (Claude Code subagents, skills, quality gates) and it self-orchestrates. Give it a 1,000 line script to follow and you slow it down.

The hidden context bill

Every session starts with a fixed cost before you say anything. Reported figures put the base Claude Code overhead around 6,300 tokens per session. Connect a single large tool server and it climbs fast: the GitHub MCP server alone is reported to inject roughly 42,000 tokens, about 21% of the 200,000 token window. (MCP, or Model Context Protocol, is the standard that lets Claude talk to outside tools like GitHub or a database.)

A bloated boilerplate CLAUDE.md stacks on top of that bill. The Code Kit's harness is built to stay out of context until a task calls for it, loading skills on demand rather than front-loading everything.

Boilerplate vs. harness: what you actually get

Competitor prices for context: ShipFast is $199, Makerkit is $299, and Supastarter starts around €349. Those products treat CLAUDE.md and agents.md as an add-on feature of the codebase. The Code Kit inverts that. The harness is the product, and the codebase is what the harness builds.

The quality gap nobody mentions

Snyk, a code security company, reported that 36% to 40% of AI-generated code contains a known vulnerability when no system enforces standards. A boilerplate with no enforcement layer ships whatever Claude generates, unchecked.

The Code Kit's agents run quality gates at every step: a type check (does the code's data line up?), a lint pass (does it follow the style and safety rules?), and a build (does it actually compile?). The database skeleton uses row-level security on every table, so one user cannot read another user's rows by default. The check is built into the harness, not left to you to remember.

What this changes

Put it together and the picture is simple. A harness with specialist agents, packaged skills, dynamic workflow support, and a CLAUDE.md under 50 lines means Claude can ship a feature in 10 to 60 minutes without you babysitting each step. The "48 hour MVP" is a harness outcome, not a boilerplate outcome. A boilerplate hands you a repo. A harness hands you a build team.

FAQ

What is a Claude Code boilerplate?

A Claude Code boilerplate is a starter codebase, usually Next.js with auth and payments, plus a CLAUDE.md file that gives Claude basic project context. You still write or prompt every feature yourself. The boilerplate gives you a starting point, not a build system.

Claude Code boilerplate vs harness: what is the difference?

A boilerplate is static code you start from. A harness is the layer of skills, agents, workflows, and a lean CLAUDE.md that tells Claude how to build. Anthropic named this distinction in June 2026. Its research found harness architecture changes output quality by 22 points and model choice by only 1.

How long should a Claude Code CLAUDE.md be?

Anthropic targets under 200 lines. Experienced harness builders target under 50. Instruction-following degrades measurably past 150 to 200 rules, and an oversized CLAUDE.md can exceed the 40,000 character file limit and eat context before any code is discussed.

What are Claude Code dynamic workflows?

Dynamic workflows shipped May 28, 2026 in version 2.1.154. They let Claude write its own JavaScript orchestration script that runs in the background, spawning up to 16 agents in parallel and up to 1,000 per run. You trigger them with the "ultracode" keyword or by asking Claude to "use a workflow."