Tooling · XiaoHu explains

Anthropic's guide to turning your manual code checks into skills that fix their own findings

Four skills run as a relay on their own team: catch bugs, clean up the diff, boot the app and look at it, and if the change touched UI, check it against the design spec.
The one-minute version
  • After the AI edits code, it can already see compiler errors, type failures, failing tests, and linter warnings. The checks only you know about are the ones you end up clicking through by hand, every single time.
  • The Claude Code team's answer: write down those steps you keep repeating, in plain language, and save them as a skill. Then it runs them itself and fixes what it finds. The smallest example is about a dozen lines.
  • Writing a good check is only half of it. The other half is which layer it lives in: you invoke it when you remember, it rides along inside the skill that generates the code, one skill calls the next automatically, or it gates every PR. Later layers take less of your attention and are harder to change.
  • Anthropic's own Claude Code team runs a four-skill relay day to day: /code-review finds bugs, /simplify cleans up the diff, /verify actually boots the app and looks at the result, and if the change touched UI they call a custom /design that checks it against DESIGN.md.
  • The costs are stated openly: once a skill loads, it stays in context for the rest of the session, so longer chains burn more tokens. And since v2.1.215, /verify and /code-review only run when you ask for them — they never fire on their own.
⚑This is the Claude Code team writing up how they use their own tool internally. The post gives no numbers: no time saved, no false-positive rate, no token overhead measured. What happens when the check itself is wrong also goes untouched. Sections marked "added by us" come from the code.claude.com docs, where the original stops short.
1What the loop looks like

Which checks the AI handles after an edit, and which still land on you

Anthropic's team published how they build verification loops in Claude Code: take the checks you do by hand and turn them into skills, so the AI corrects itself. Written by Delba de Oliveira of the Claude Code team, published July 22.

A verification loop just means Claude checks its own work after writing code and fixes what it finds — no one has to sit at the keyboard supervising.

Here's roughly how it goes today. You ask the AI to change a feature, it reports back that it's done, and then you open the page and click through it, glance at the logs, and drag the browser window down to phone width. Those few motions are on you every time. Those are what you're handing over.

Start with the shape of the full loop.

Anthropic diagram: the agentic loop from prompt through gathering context, taking action, and verifying results, with a dashed box around the agentic loop
You give it a prompt; Claude gathers context, takes action, verifies results, and reports back. The dashed box holds the three middle steps — the loop it runs on its own. Source: Claude's official blog.

One detail in this diagram the post doesn't mention: when verification fails, the arrow points back to "gather context," not to "take action." It doesn't retry the same edit with the same information — it goes back and re-reads the situation before deciding what to do next.

It already handles part of the verification step, because a codebase is full of signals a machine can read directly: compiler errors, type failures, failing tests, linter warnings. The moment one appears, it sees it and knows to circle back.

The other half is the problem.

It sees these
Deterministic signals in the codebase, each with an error message attached
  • Compiler errors
  • Type check failures
  • Failing tests
  • Linter warnings
Only you know these
Nothing throws an error for them, so it can't infer them
  • Whether this button gets clipped at phone width
  • Whether this error log dumps the user's request body
  • Whether the migration that dropped a column backfilled the data
  • Whether this change follows your team's design spec

Nothing in the right-hand column raises a flag for you. So they fall back to you every time, as those few manual clicks. The job now is to turn them into signals it can read too.

Related on this site
Claude Code's official guide to loop design: four levels up to unattended
The previous post in the same official series, on how a loop goes from manual confirmation all the way to running unattended. Verification is one of those four levels; this piece is where it opens up.
2What's already there

The six checks that ship with Claude Code, and what each covers

Before writing anything from scratch, look at what's already built in. Claude Code ships six, each sitting at a different point in the loop — some you invoke, some run on their own.