A second review by another AI

This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.

By the end of this lesson, you will be able to

choose a second pair of eyes: new conversation, another model, another assistant;
give the reviewer what it needs, and no more;
get a list of problems, with evidence and a level of certainty;
decide each point yourself.

Why it matters to you

A tool that produced a text does not always see its own mistakes. A second pair of eyes finds others. But it sometimes makes some up too, with the same confidence. A cross-review is therefore not “the second AI is right”: it is a list of leads, which you check against the source before deciding.

Before you start


You need	A second pair of eyes (a new conversation, another model or another assistant), your brief and the result from lesson 3, and the review instruction (resource)
Cost	€0 with a new conversation, or a second assistant on a free account. Codex is included in the ChatGPT plans, including Free, with limits; for Free and Go, the Codex plans page only mentions the desktop app (25/10/2026)
Duration	35 minutes
Reminder	The reviewer receives the brief, the source and the result: nothing from the “never” list

Understand it in 3 minutes

Three levels of second pair of eyes, from the simplest to the most different:

Second pair of eyes	How	Difference from the first
A new conversation, same tool	New conversation with no memory (incognito chat or memory turned off, lesson 2), you paste the brief, the source, the result	No production history
Another model, same tool	In Claude Code: a new session started with `--model` (for example `claude --model sonnet`), or `/clear` to start from zero, then `/model`	Another model, which does not necessarily make the same mistakes
Another assistant	ChatGPT, Claude, Codex… by copy-paste, or through a real integration (lesson 5)	Often another company (Claude and Claude Code: Anthropic; ChatGPT and Codex: OpenAI), so other models

What you give the reviewer: the brief (the request), the source, and the result. Not the production conversation, nor the first tool’s explanations.

What you ask of it: a ranked list, from most serious to least serious; for each point, the evidence (an exact quote) and the certainty (sure, probable, to be checked); no rewriting; no file modified. The full instruction is in the resource.

What you do next: for each point, you check against the source, then you decide: correct it, ignore it (noting why), or talk to someone about it. One round trip, two at most (lesson 7).

If you use Codex (OpenAI’s tool that reads, edits and runs code, included in the ChatGPT plans): according to its documentation (24/10/2026), its /review command reviews code changes in a project and returns a list of ranked problems, without modifying your files. It is made for code, not for reviewing a text like the invitation: for the invitation, paste the instruction, the brief, the source and the result as in step 2.

OpenAI also publishes a plugin for Claude Code, whose /codex:review command does a read-only review of the changes in your project; it does not take a custom instruction, and using it counts toward your Codex limits.

According to the Codex plans page (25/10/2026), Free and Go give access to Codex in the desktop app, subject to rollout; the terminal, which the plugin for Claude Code uses, and the local /review command are only listed from Plus. The plugin’s documentation nevertheless says “including Free”: check with your account.

Do it together

1Step 1: Prepare the review

Go back to the club exercise (lesson 3): the brief, the notes, and the invitation obtained with the brief. Open the resource Handoff brief and review instruction, part 2.

2Step 2: Run the review

Open a second pair of eyes: a new conversation, another model, or your second assistant. Paste the review instruction, then the brief, the notes and the invitation, each with a heading (“THE BRIEF”, “THE SOURCE”, “THE RESULT”).

In our trial of 24/10/2026, the review by a new Claude Code session, with another model (Sonnet), gave:

a real problem, which the production session had copied without seeing it: “November 22 is a Saturday in 2025, but a Sunday in 2026”, marked “to be checked”;
a useful point on form: “Markdown bold will not come through in a plain-text email”;
a false problem: it suspected the AI of having created the invitation file, when it was us. It had marked it “to be checked”, rightly.

3Step 3: Decide

Fill in the table in part 3 of the resource. For the exercise:

the date: check in a calendar. In 2026, November 22 is a Sunday. Decision: ask the club which day the room is really booked for, before sending;
the bold: correct it, if your email is plain text;
the file: ignore it, because you know who created it.

4Step 4: Wrap up

Correct the result yourself, or ask the production tool for a single targeted correction (“Fix only these two points: …”). Then stop: the review has done its job.

Your turn

Have a second pair of eyes review the result of your task (lesson 3), with the review instruction.
Check each point against the source, and fill in the table for deciding.
Write in your journal: the reviewer used, the true points, the false points, and your decision for each.

Check your work

It worked if:

the reviewer received the brief, the source and the result, and nothing else;
you got a list of problems, with evidence and certainty levels, without a rewrite;
each point was checked against the source, then decided by you;
you stopped after one or two round trips.

Common mistakes:

applying everything the reviewer says, without checking;
giving the reviewer the production conversation: it then judges the intentions, not the result;
asking for a complete rewrite, and having to read everything again;
having the review reviewed, then the review of the review.

Keep it and pick it up later

Keep in your journal, for each review, the list of points and your decisions.
If a reviewer is often wrong on the same type of point, note it: it is useful information for choosing your tools (lesson 2).

Going further

optional

The Codex plugin for Claude Code is published by OpenAI: github.com/openai/codex-plugin-cc. Before installing it, apply the method from module 7, lesson 6 (provenance, permissions, cost), and check its prerequisites: Claude Code (which is not included in Claude’s free plan), Node.js 18.18 or later, and Codex installed on your computer and signed in with a ChatGPT account or an API key.

New words

4 words explained

Word	What it means
Cross-review	Having a tool’s work checked by another point of view
Second pair of eyes	A new conversation, another model, or another assistant
Deciding (point by point)	Choosing, point by point, what you correct and what you ignore
Read-only	The reviewer reads, but changes nothing

The review in step 2 was really run with Claude Code 2.1.310, Sonnet model, read-only (test report, T45). To be checked: the review by Codex and by ChatGPT has not been tested for this lesson.