First-party. Harness Teardown is built and maintained by ShipWithAI. Unlike the reviewed third-party tools in this toolkit it is our own product, and it carries no review score — judge it by the findings it gives you.

→ Run the free teardown — no install, no signup.

Problem it solves

Everyone’s second question about a Claude Code harness is “is mine any good?”, and there has never been a cheap way to answer it. Reading someone else’s CLAUDE.md tells you what they decided, not what you got wrong. So harnesses drift: a permissions block nobody has reviewed since the first week, hooks that reference scripts that moved, a memory directory with four files that contradict each other.

Harness Teardown turns that into a number plus a list. You paste the files, it runs seven deterministic check families over them, and it hands back a static score with each finding pinned to a line and a suggested fix.

How to use

  1. Open /harness-teardown/.
  2. Pick your project folder, or add the files by hand. CLAUDE.md is required; the rest are optional and each one it finds sharpens the result.
  3. Press Run static check. Nothing leaves the page before that press.
  4. Read the score band and work down the findings. Each one names the file, the line and the fix.

The optional AI deep analysis is a second, separate step — the free static score never depends on it.

Pro tips

  • Run it before a rewrite, not after. The usual result is that CLAUDE.md was fine and the permissions or hooks were the real problem.
  • Re-run after each change and watch the band move. Because the checks are deterministic, the delta is attributable to your edit.
  • Treat the score the way you treat test coverage: useful as a trend, dishonest as a target.
  • Want the same checks in your terminal? Harness Doctor runs them inside Claude Code, so CI and the web page agree.

When NOT to use

Skip it if you have no harness yet — there is nothing to score, and ShipWithAI Harness or ShipWithAI Starter is the better first stop. Skip it too if what you actually want is a judgement on your instructions rather than your structure: static checks cannot tell you that a rule is bad advice, only that it is missing, duplicated or unreachable.