Making it attack your work instead of writing it

Turn the tool around

Almost everything so far has the model producing text you then check. Reverse it: you write, and the model attacks.

This is the safest mode there is, for a mechanical reason. You supplied the entire text, so the material it works on is all in front of it and all verifiable by you. There is very little for it to invent, and anything it does invent you can catch by reading your own document. The failure mode that dominates the rest of the course is largely absent here.

It is also the mode that most improves work you were already going to produce, which makes it the easiest thing to defend to a sceptical colleague.

Never ask whether it is good

Ask "is this any good?" and you will be told it is strong, clear and well-structured, with one gentle suggestion. This is sycophancy, and it comes from the same preference-tuning that produces the hedging in the previous lesson: agreement is rated highly by human raters, so agreement is what the training pushed towards.

The way past it is not to plead for honesty. It is to remove the yes/no question entirely and demand a fixed quantity:

Give exactly five objections to this document, ordered from most to least serious. For each, quote the sentence that causes the problem.

Asking for exactly five forces it past the first agreeable answer. Objections one and two are usually obvious, three and four are often the useful ones, and five is frequently filler — which is fine, because you asked for five and you can discard one.

The prompts that earn their keep

The hostile reader. "List every factual claim a hostile reader would demand evidence for. Quote each claim. Do not evaluate whether the claim is true." This finds the assertions you made from memory and never sourced.

The missing question. "You are the finance director reading this for the first time. What is the first question you ask, and where in the document should the answer have been?" Naming the reader's role is doing real work here — it is a description of what that reader cares about, not a costume.

The internal contradiction. "Find every place a number, date or name appears more than once. Quote each occurrence and say whether they agree." This is dull, mechanical, and something models are genuinely good at. It catches the figure updated in the summary and not in the appendix, which is the error that most often reaches a board paper.

The defined-term check. "List every term this document defines or uses as if defined. For each, quote every place it is used in a different sense." Contracts, policies and specifications rot in exactly this way.

The steelman. "Write the strongest three-paragraph objection to this proposal from the operations team's point of view." Read it before the meeting rather than hearing it in the meeting.

The regret test. "Which single sentence in this document would be most damaging if it were read aloud in a dispute, and why?" Blunt, and it works.

Make it quote, then check the quote

Every one of those prompts asks for a quoted span. Insist on it, and then confirm the span really appears in your document — a search in your word processor takes two seconds.

This is not pedantry. A critic that paraphrases can criticise something you did not write, and you will spend ten minutes fixing an imaginary problem. Requiring a quote turns each objection into a claim you can verify instantly, and the rare fabricated quote tells you to treat the rest of that reply with more suspicion.

What it cannot see

Be clear about the ceiling, because it is a real one.

The model does not know that the operations director has opposed this twice before, that the finance team is defending a budget line, or that the phrase you used in paragraph two is the phrase somebody was sacked over. Its objections are generically strong, not specifically strong. They are a floor: the things any competent reader would raise.

The specific objections — the political ones, the historical ones, the personal ones — remain yours to anticipate, and they are usually the ones that decide whether a proposal survives. Treat the generated list as the part of the review you no longer have to do by hand, and spend the time you saved on the part it cannot do.

It is also the wrong tool for judging whether an argument is correct in a field you do not know. It will find that your reasoning is unsupported. It will not reliably find that your reasoning is wrong.

The confidentiality note, briefly

This mode involves pasting your document into something. Everything in the later block on confidentiality applies in full, and it applies most sharply here, because the documents worth criticising are usually the sensitive ones.

The good news is that criticism is a task small models do respectably. A seven- or eight-billion-parameter model running locally through Ollama or LM Studio, both free, will find contradictions, unsupported claims and inconsistent terms in a document that never leaves your machine. It will be less incisive than a frontier model. On this particular job, the gap is smaller than you would expect, and for a document that may not leave the building it is the only version of the option that exists.