I review everything Claude Code writes, but at the current speed increase, it's easier to miss things. So after my review, a second, independent reviewer checks the work, and it gets more than the diff.

Here's how I've set it up.

Claude Opus generates, I review, Codex does a second review

The idea here is to use a different model/tool (in my case Codex) to keep bias out. But instead of a pure diff, which is already helpful, the reviewer also needs to know what the aim was.

For serious work (one-off scripts can get away with less), I start each feature with a grillme session to come up with a concise and well-scoped spec. I updated my variant of this skill to write the resulting spec to a local specs folder (goal, decisions (including rejected options), in/out of scope, acceptance). This way the reviewer has much richer context about the diff it's reviewing.

This is the spec template my grillme skill writes to specs/<slug>.md, kept under ~40 lines:

# <Feature>

## Goal
## Decisions
## Scope
## Acceptance
## Out of scope

If you're new to skills, check out how I use Claude Code skills for structured AI collaboration.

This was also a succinct comment from somebody on X when I posted about this last week:

"a reviewer with zero context from the writing session, otherwise it's grading its own homework. Do you give the reviewer the spec, or only the diff?"

So the answer is both!

A tight CLAUDE.md (touch only what the task needs, grep callers before changing a signature, ruff → ty → pytest) stops the simple mistakes before review.

So my flow for more serious work: grillme → spec → Opus implements → I review (!), then Codex for a second, independent and well-scoped review.

Recent example: Codex raised one P2. Each input passed validation on its own, but their product overflowed the DB column. Unlikely in practice for that app, but cross-field constraints like this are easy to miss (I've seen it in 1:1 coaching more than once).

The diff tells the reviewer what changed. The spec tells it whether that was the point.