You will be able to get a useful critique from an assistant while knowing what self-checking cannot catch.
Ravi had written a one-page proposal for his cafe group's management: a loyalty app to replace the paper stamp cards. He had drafted it with an assistant over an afternoon, and at the end he asked the same chat, "Is this good?" The reply praised the structure, called the argument persuasive and suggested adding a call to action. He sent it. In the meeting, the finance director asked how much the app would cost to run each year. The proposal did not say. Neither did the critique.
Asking an assistant to critique work is useful, but only if you ask in the right way, and only if you know what it cannot catch.
"Is this good?" invites a polite answer, and AI fundamentals lesson 4.3, Human feedback and why assistants sound so agreeable, explained why assistants lean towards telling you what you would like to hear. A general request for feedback tends to get general praise and a few small suggestions.
Specific questions get specific answers. Three work well for almost any piece of work:
What are the three weakest points in this, and why? Which claims are most likely to be wrong or need checking? What would a sceptical reader, such as a finance director who has to approve the budget, object to?
The first forces a ranking, so the model cannot say everything is fine. The second points it at facts rather than style. The third puts it in the shoes of the actual reader, which is where Ravi's missing running cost would have come up. Naming who the sceptical reader is makes the third question much sharper. A finance director, a cautious client and a new customer object to different things.
Critique works better in a fresh chat, without the history that produced the draft. A chat that has just spent an afternoon writing a proposal with you has seen every reason you gave for each choice, and those reasons are still in view, so it tends to agree with them. It is also critiquing text that grew out of its own earlier messages, which pulls it towards defending what is there.
A fresh chat sees only the draft, the way your reader will. Paste in the finished piece and a line of context about who it is for, then ask the three questions. You can also strip out anything that signals how much work went into it. You want the reaction of someone reading it cold.
Ravi tried this afterwards. In a fresh chat, the three questions produced, among other things: "The proposal gives the one-off development cost but no annual running cost, such as hosting, maintenance or app store fees. A finance reader will ask for it." That was the question he had faced in the meeting.
There is a limit to all this, and it matters. A model checking its own kind of output shares its own blind spots. If the draft contains a fact the model invented with confidence, such as a statistic about loyalty app adoption that sounds plausible, the same model reading it later is likely to find it plausible too. AI fundamentals lesson 6.1, Why a fluent answer can still be false, explained that there is no built-in step that checks claims against a source. Asking for a critique does not add one.
So a critique is good at catching gaps in reasoning, missing sections, unclear structure, weak arguments and things a reader would ask about. It is much weaker at catching errors of fact, especially specific ones that sound right. The second question, about claims most likely to be wrong, helps because it gives you a list to check. The checking itself is still your job, using the methods in lesson 7.3, Checking numbers, quotes, sources and code.
You can also paste the draft into a different assistant and ask the same questions. This sometimes catches things the first one missed, because different models have been trained differently and have different weak spots. It is a cheap, quick extra check for work that matters.
Be careful what you conclude, though. When two assistants agree that a claim is correct, that is not proof. They may have learnt the same wrong fact from similar sources, or both may be filling a gap in the same plausible way. Agreement between two models is weaker evidence than one check against a primary source, and it should never replace one when the stakes are high, as lesson 7.1, Match the check to what is at stake, set out.
The useful way to think about a critique is as a reader who never gets tired and never gets offended, but who cannot look anything up for you. To see what that reader is worth on your own work, take a draft you have finished recently, ideally one you produced with an assistant, and open a fresh chat to put it in.
Paste a draft into a fresh chat, ask for a critique using the three questions, and record which points were useful and which it missed.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).