Self-Critique: Self-Refine, Chain of Verification and Pre-Mortems

Study guide

Make AI Review Its Own Work

Most people treat the first answer from ChatGPT, Claude or Gemini as the finished product. It is not. It is a starting point, and it often contains gaps, weak arguments and confident errors that sound completely plausible. This course teaches you six established techniques for making the AI review its own work before you send, publish or decide anything. You will learn self-refine, chain of verification, constitutional critique, self-reflection rubrics, pre-mortems and devil's advocate prompts. Each one is a short, practical step you can add to the work you already do.

This course is for busy professionals who care about quality and accuracy more than speed. If you write proposals, reports, emails or summaries, or if you make decisions you cannot easily reverse, these methods will save you from embarrassing mistakes. You do not need any technical background. You need one extra message and a minute of reading. By the end, you will have a repeatable habit: attach one review prompt to every task that carries real weight, choose the lens that matches the risk, and keep your own judgement as the final check.

What is a prompt framework?

A prompt is the request you type into an AI tool such as ChatGPT, Claude or Gemini. A prompt framework is a reusable structure for that request: a short checklist of the parts a good prompt contains, usually named with an acronym so it is easy to remember. Why it helps: the AI fills in whatever you leave out with generic guesses. A framework makes sure you include the parts that change the answer, such as who it is for, what you want to achieve and the format you need.

Frameworks make prompts repeatable: you fill the same parts each time, save the prompt as a template and share it with colleagues. They are not magic words: the value comes from the detail you put in each part. They help most when a request is complex or open to interpretation; a quick, simple question can stay short. There are well over sixty named frameworks, and most of them remix the same ingredients: role, task, context, audience, format, constraints and examples.

In this course you learn Self-critique, which is a prompting technique from AI research: instead of slots to fill, it changes how the AI works through a task, and you can add it on top of any fill-in framework. It is best for: anything where quality and accuracy matter more than speed.

The first draft is not the answer

AI models are stronger at spotting problems in existing text than at producing a perfect first version. That is the core insight behind self-critique. A review pass catches gaps, weak arguments and unclear structure that the first pass missed. The Prompt Report, a survey of prompting techniques, lists self-critique as one of six established families of prompting. Anthropic describes draft, review and refine as the most common prompt chaining pattern. It is a sequence, not a single clever prompt, so plan for two or three messages.

A good critique prompt names the criteria, such as clarity, structure and weak arguments. Vague instructions like make it better give vague revisions. The review step costs you one extra message and a few seconds. The return is a draft you can send or publish with far less editing by hand. Think of it as a second pair of eyes that never gets tired and never takes offence.

Self-refine: draft, critique, revise

Self-refine comes from Madaan and colleagues, 2023. It loops three moves: draft, critique the draft, revise using the critique. No human editing sits between the steps. A critic persona sharpens the feedback. Asking for a critique as a demanding editor or a sceptical CFO sets a standard the model can aim at, instead of polite generic comments. The persona you pick should match the risk in the document. An editor for clarity, a CFO for cost and return.

You control the revision. Ask the model to revise using only the critique points you agree with, so your judgement stays in the loop and good lines are not rewritten. Splitting the loop into turns helps. Ask for the critique first, read it, then ask for the revision. Asking for both at once often produces a defence of the draft rather than an improvement. The loop works on any text task: proposals, reports, emails, summaries.

Chain of verification: check every claim

Chain of verification was described by Dhuliawala and colleagues in 2023. It has three steps: list every factual claim, check each against the source, then rewrite with corrections. The prompt is simple: list every factual claim you made, check each against the source, rewrite with corrections. Paste it after any AI answer where facts matter. This method is one of the strongest everyday defences against confident errors. AI models are trained to sound sure, not to be right, so a separate check step is essential.

Always provide the source when you ask for verification. Without a source, the AI will simply agree with its own previous answer and the check is useless. Check every claim, not just the ones you doubt. Errors often hide in details that sound plausible, like a date or a name that feels familiar. The rewrite step is not optional. If you skip it, you still have the original errors in your document. The correction only happens when the AI produces a new version.

Critique against principles and rubrics

Constitutional critique, from Anthropic in 2022, reviews output against written principles rather than gut feel. In practice your principles are your style guide, tone rules or company policy. OpenAI's self-reflection rubric asks the model to build a five to seven category rubric for excellent work first, then check the draft against it. You approve the rubric before the review runs. A rubric turns make it better into specific, repeatable checks. Each category names a quality, and each score points at a line in the draft.

Always ask for evidence, meaning the quoted line that supports each score. If the model cannot quote a line, the finding is probably invented. Keep the number of categories between five and seven. Fewer makes the review shallow, more makes it drift and lose focus. Reviewing against written principles makes results repeatable across drafts, people and weeks, which is what a compliance or quality process needs.

Pre-mortem and devil's advocate

A pre-mortem assumes the plan already failed six months from now and asks what went wrong. Working backwards from failure makes it easier to name risks that a forward-looking review would skip. A devil's advocate prompt asks for the strongest case against a decision, not a balanced view. Balance softens the counter-argument, while the strongest case forces it to be built properly. Both techniques are cheap: one message and about a minute of reading, compared with finding out months later when the money is already spent.

Ranking the failure causes, most likely first, turns a vague worry list into something you can act on. Add a request for the earliest warning sign of each risk. The most valuable output is usually the quiet risk, the thing people believe but will not say out loud in a planning meeting. A pre-mortem gives them a safe way to surface it. Run a stress test before any decision you would struggle to reverse. For reversible decisions, the cost of checking may outweigh the benefit.

Building a review step into your job

A review prompt is a short instruction that asks the model to critique, verify or pre-mortem a draft. Attach it to every task that carries real weight, not just the big ones. Match the lens to the risk. Numbers for anything financial, facts for claims and names, tone for anything a customer reads, legal for wording that creates commitments. Ask for problems only, not a rewrite. This keeps the draft yours and stops the model from quietly changing your meaning or voice.

The model reviews, you decide. Treat every note as a suggestion. If a flag is wrong, ignore it. If the model says something is fine, that is not proof it is fine. A review you do not act on is wasted effort. Read the notes, sort real issues from noise, and make one or two concrete changes before you send. The whole course points to one habit. Self refine, chain of verification, principles, pre mortems and devil's advocate are all ways of building a review step into your work.

Frequently asked questions