A second model is not a luxury, but quality work

A persuasive AI answer can easily look as though it has already been checked. It is fluent, neatly structured and sounds decisive. That is precisely what makes it dangerously convenient. Quality does not come from a draft sounding confident. It comes from someone — or something — trying to disprove it.

A draft is not yet a judgement

Anyone using AI for an important task needs two distinct movements: create something, then examine it. The first step can be fast, curious and generous. The second must work differently. It looks for gaps, unsupported leaps, missing perspectives and sentences that claim more certainty than the available basis permits.

A person can do this. A second model can do it too — with a deliberately different brief. The key is not simply the number of tools. The key is that the second look does not repeat the job of the first.

Independence begins with a different question

A second model becomes useful when you do not ask it for agreement. Do not instruct it to make the text prettier. Ask it to put the draft under pressure: Which claim is too strong? Which condition has been assumed silently? Which statement cannot stand without a source? Who might misunderstand the text, or have good reason to object?

Those questions change the character of the collaboration. The second model is not a copy editor; it is an adversarial reviewer: respectful, specific and willing to disturb a good feeling. Its task is not to be right. Its task is to make the cost of being wrong visible.

The hand-off determines the quality of the review

A review without material stays generic. Pass along not only the outcome, but also the goal, audience, permitted sources and the decision that depends on it. Say explicitly what the second model must not speculate about. Ask it to identify locations in the draft rather than offering vague criticism.

A good review brief also limits scope. Three to five prioritised findings are usually more useful than twenty small remarks. Every finding should include a place in the draft, a reason and a next action: support it, qualify it, rephrase it or remove it.

Not every difference is an error

Two models will not always say the same thing. That is not a defect; it is information. When a review disagrees, do not automatically adopt the newer answer. Check whether the criticism identifies a real gap, whether it can be grounded in your material and how consequential the point is for the decision.

Responsibility remains with you. The second model offers a reason to think, not a release to publish. That is why this sequence is quality work: it turns a smooth individual output into a traceable decision process.

Every usable finding leads to exactly one of four decisions:

An example: the robust recommendation

Suppose you recommend a tool to a management team. The first draft sounds convincing: clear benefits, a price, a timeline. Before it goes out, you give it to a second model — not to polish it, but with the brief to attack it.

The review reports three findings: the stated saving rests on a single source; a compatibility assumption was never made explicit; one sentence promises certainty the contract does not cover. For each it names a place, a risk and a possible step.

Now you sort: the saving is supported or qualified, the assumption is made visible, the over-strong sentence is rephrased. None of this weakens the recommendation. It becomes more robust — and later you can show what each statement rests on.

A review brief for the second look

# REVIEW BRIEF

**Purpose of the draft**
It should …

**Material that applies**
Use only …

**Check especially for**
Unsupported claims, silent assumptions and missing perspectives.

**For each finding, provide**
1. Location in the draft
2. Risk or reason
3. Next action: support, qualify, rephrase or remove

**Boundary**
Do not write a new version. Give at most five prioritised findings.

A second model does not make a draft automatically true. It makes the path of verification more visible. When an answer withstands an independent counter-question, not only the text improves. Your decision becomes more robust too.

Worksheet: Let a draft be challenged deliberately

Choose a draft whose content affects a real decision, publication or recommendation. Run a short, clearly bounded review.

Mark the draft. Mark three claims that matter most to the outcome. Note what supports each one.

Give the review brief. Pass on the draft, goal and material. Ask for no more than five concrete findings with location, risk and next action.

Sort the findings. Assign every finding to one of four decisions: support, qualify, rephrase or remove.

Record the decision. Write down which change you adopt — and why. Also record one criticism you reject with a reason.

Both working materials for this article — the topic overview and the worksheet with a reflection space to fill in — are available for download here: