Why this lesson exists. Lessons 2 and 3 used the model to produce things. It is at least as useful for finding what is wrong with things, including what is wrong with its own output. This lesson closes Part A by turning the model from a writer into a second pair of eyes, which is the habit that makes everything later in the course trustworthy.
By the end of this lesson you will be able to use the model to critique, verify and stress-test work, because you will have run a critique on something real and seen what it catches.
A founder has spent two days on a proposal for a large prospect. It is good. Before sending, she pastes it in with this:
Read this proposal as the prospect's finance director, who is sceptical of software vendors and has been burned by a failed implementation before. List every objection you would raise, in order of how much each would worry you. Do not soften them.
The model comes back with seven objections. Five she had already addressed. One is a phrasing issue. The seventh points out that the proposal promises a four-week build but the pricing section describes a twelve-month contract with no mention of what happens if the build is late, and a finance director who has been burned before will read that gap as the whole risk of the deal. She adds two sentences on delivery guarantees. The deal closes, and the finance director mentions in the kick-off that the delivery clause was what got it over the line.
The model is a poor judge of its own confidence (ask it "are you sure?" and it will say yes) but a strong critic when given a role and a target, because criticism is pattern-matching against everything it has read about what goes wrong. It never gets tired, never worries about offending you, and can be told to adopt exactly the point of view you are worried about.
Three kinds of check are worth making habitual:
Critique. "Read this as [the person most likely to object] and list every objection." Works on proposals, emails, plans, pricing, slide decks.
Verification. "Here is the report and here is the source spreadsheet. Check every figure in the report against the source and list any that do not match." The model reads both and compares, which is exactly the tedious job humans skip.
Stress-testing your own view. "I think the real objection is price, not migration. Argue the other side as strongly as you can." The model is a good sparring partner and a poor mind reader; tell it what you think and ask it to attack it.
The reason to make this a habit rather than an occasional trick is that it changes where your attention goes. Instead of re-reading your own work looking for problems (which you are bad at, because you wrote it), you read a list of candidate problems and decide which are real. That is a better use of a human.
Everything the model produces can be checked the same way. When it summarises a document, ask: "List three claims in your summary and quote the sentence in the source that supports each." When it analyses numbers, ask it to show which rows it used. When it gives you a fact you will repeat out loud, ask for the source and go and look. This is how you use the model's speed without inheriting its confident guesses, and it is the foundation for trusting agents later on, when you will not be watching.
