What the diff doesn't show

Suppose you change how an application decides whether an item is available. The product page follows the new rule. Checkout still uses the old one. The diff looks sensible because checkout never appears in it.

That’s the kind of gap I want a coding agent to look for before calling the work finished.

The changed files tell a reviewer what was edited. They don’t identify every place that should have changed. Finding those places needs to be part of the task.

Skills give me somewhere to describe where the agent should look and what evidence it should bring back. I’m organizing those instructions into a workflow I can reuse across projects.

Before editing, the agent needs the repository’s CLAUDE.md or AGENTS.md and the closest existing implementation. reuse-first directs that search. A small plan should name the affected behavior and the check that would expose a mistake.

After building, completeness searches for other places that read the same value or make the same decision. In the availability example, that means finding checkout. Each relevant place needs an answer: changed, or left alone for a stated reason.

A successful response to an order request doesn’t prove that stock was reserved correctly. verify asks for evidence appropriate to the change: perhaps a targeted test, a real request, or a query of the stored state. I want the smallest check that could actually expose the fault.

For changes involving authentication, money, tenant isolation, concurrency, or data loss, I want a separate code review after those checks. In Codex CLI, codex review --uncommitted reviews the local changes; the app also offers /review. Current Claude Code includes /code-review for reviewing a local diff. These can run before opening a pull request. Codex review, Claude Code review.

A review finding still needs checking against the implementation. If it reveals a defect, the fix needs its own completeness and verification checks. A clean review doesn’t make those earlier steps redundant.

Using the workflow

The source repository contains reuse-first, completeness, and verify, with separate Claude Code and Codex versions. They share the workflow but use each tool’s repository instructions and review commands. Copy the complete folders from claude/ into your project’s .claude/skills/, or from codex/ into .agents/skills/. The README includes installation commands for each version.

Ask the agent to use reuse-first before editing, then completeness and verify after building. Put that order and the conditions for a separate code review in your project’s CLAUDE.md or AGENTS.md. Installing the skills makes them available; those project instructions describe how they fit together. The review itself uses the tool’s reviewer.

That’s the handoff I want from an agent: the places it checked, the behavior it observed, and the assumptions that still need attention. With that in front of me, I can make a decision about the change.