Critique-First Prompting: Make AI Judge Before It Writes
Critique-First Prompting: Ask an expert Callstand to judge the brief or the existing draft before asking any model to write, so generation starts from known problems. It applies the "LLM as a judge" idea at the start of the work instead of the end. The judge finds what's wrong with your brief or current draft, and that list becomes the constraints for whatever gets written next.
Most AI feedback happens too late. You ask for a landing page, get a competent average one, then spend three rounds of "make it punchier" trying to steer it somewhere better. Each round fixes the surface and leaves the real problems in place: a brief that never said who the page is for, an offer nobody would believe, a structure that hides the answer.
Critique-First Prompting reverses the order. Before any model writes, you have it judge what you already have, whether that's the brief, last quarter's version or a teammate's draft, against a real standard. Generation then starts from a known list of problems rather than a blank page. This guide covers the pattern, the LLM-as-a-judge research behind it, copyable prompts, and how to keep an AI judge honest.
- What is Critique-First Prompting?
- What is LLM as a judge?
- Why judging first produces better writing
- How to run Critique-First Prompting in four steps
- Critique-first prompt templates
- Example: AI feedback on a landing page draft
- Critique-first vs other prompting patterns
- How to keep an AI judge honest
- Running it with Callstand
- FAQ
What is Critique-First Prompting?
Critique-First Prompting: Ask an expert Callstand to judge the brief or the existing draft before asking any model to write, so generation starts from known problems.
Three parts of that definition do the work:
- An expert judges, not a generic model. The critique uses someone's actual standard for what good looks like. Without one, you get the model's average opinion, which is the thing you're trying to get away from.
- The brief or the existing draft gets judged. Most writing tasks don't start from nothing. There's a brief, an old version, a competitor's page or a rough draft. Those are where the problems already are.
- Generation starts from known problems. The critique isn't the end product. It becomes a short list of constraints the writing prompt has to satisfy.
It's a cousin of the SDC Loop (Strategize → Draft → Critique). The SDC Loop puts critique at the end of a cycle. Critique-First Prompting puts it at the front, for the common case where you're improving something that exists rather than starting fresh.
What is LLM as a judge?
LLM-as-a-judge is the practice of using a large language model to evaluate output, such as scoring an answer, picking the better of two responses or checking a draft against criteria, instead of asking a person to do it. AI labs started using it because human evaluation is slow and expensive, and the research on it is what makes critique-first work.
The best-known study is "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" (Zheng et al., 2023). It found that strong judges like GPT-4 agreed with both controlled and crowdsourced human preferences over 80% of the time, "the same level of agreement between humans." The same paper documented the judge's weak spots: position bias, verbosity bias, self-enhancement bias and limited reasoning ability.
Other work showed how much the setup matters. G-Eval (Liu et al., 2023) had GPT-4 reason step by step and fill in a structured form for each criterion, which improved its agreement with human ratings of summaries over earlier automatic metrics. It also flagged that LLM evaluators may favor LLM-generated text. And "Large Language Models are not Fair Evaluators" (Wang et al., 2023) showed that simply swapping the order of two answers could flip results: with ChatGPT as the judge, Vicuna-13B could beat ChatGPT on 66 of 80 test queries.
The takeaway for everyday work: an AI judge is good at applying criteria you give it and unreliable when it has to invent them. That's why Critique-First Prompting starts from an expert's standard.
Why judging first produces better writing
The research on AI feedback points the same way from three directions.
- Feedback improves output. Self-Refine (Madaan et al., 2023) had a model draft, critique its own draft and revise. Across seven tasks, people and automatic metrics preferred the refined outputs, an improvement of about 20% absolute on average over one-step generation.
- The feedback needs an outside reference. "Large Language Models Cannot Self-Correct Reasoning Yet" (Huang et al., 2023) found that without external feedback, models struggled to correct their own reasoning, and sometimes got worse after trying. A critique grounded in someone else's standard is that external reference.
- Written principles make critique usable. Anthropic's Constitutional AI (Bai et al., 2022) trained a model by having it critique and revise its own responses against a written list of principles. Critique tied to explicit rules is something a model can act on.
Put together: critique helps, critique against a real standard helps more, and critique works best when the standard is written down. Doing it before generation adds one more benefit. The writing prompt is shaped by the problems from the start, so you skip the rounds of fixing a draft that went wrong in its first paragraph.
There's also a cost argument. Every "make it better" round is context you type again, which is how Prompt Debt builds up. A critique that names the problems once replaces a lot of retyping.
How to run Critique-First Prompting in four steps
1. Pick what gets judged
Choose the input closest to the final piece: the brief, the current live version, or a rough draft. If you only have an idea, write a five-line brief first (audience, goal, offer, constraints, what "done" looks like) and judge that.
2. Judge it against a standard
Ask for problems, not a rewrite. Name the criteria, or use a connector that already carries them. Ask for three groups: what to fix, what's fine, and what to stop worrying about. The last group matters: it stops the writer from "fixing" things that weren't broken.
3. Turn the critique into constraints
Rank the problems and keep the top three to five. Rewrite each as an instruction: "the page never says who it's for" becomes "open by naming the reader: finance leads at 50 to 500-person companies." This short list is the known problems the next step starts from.
4. Write, then check against the same list
Now ask for the draft, with the constraint list at the top of the prompt. When it comes back, check it against the same list. Using the same standard before and after keeps the judgment consistent, which is the point the Judgment Layer makes about where output quality actually comes from.
Critique-first prompt templates
Copy these into any AI tool. Replace the bracketed parts.
Critique the brief
Before writing anything, judge this brief as [role, e.g. a senior B2B copywriter].
Criteria: [3-6 criteria, e.g. clear reader, one believable promise, proof available, a single CTA].
Return three lists: what to fix (most important first), what's fine, and what to stop worrying about.
Do not write the piece.
Brief:
[paste brief]Critique the existing draft
Judge this draft against the criteria below. For each criterion, quote the passage you're judging,
then say pass or fix and why. Evidence first, verdict second. Do not rewrite it.
Criteria: [criteria]
Draft:
[paste draft]Write from known problems
Write [deliverable] for [reader].
Known problems to solve (from the critique):
1. [constraint]
2. [constraint]
3. [constraint]
Keep what the critique marked as fine. Ignore what it said not to worry about.
When done, list how the draft handles each known problem.The "evidence first, verdict second" line comes straight from the bias research: Wang et al. found that asking the judge to produce evidence before a rating made evaluations more reliable.
Example: AI feedback on a landing page draft
An illustrative example, not a client result. A team has a landing page draft for an invoicing tool and wants AI feedback on the writing before a rewrite.
Write-first approach: "Improve this landing page." The model tightens sentences, adds a few power words and moves the CTA up. The page reads more smoothly and still doesn't work.
Critique-first approach: the draft goes to a judge with the team's criteria first. The critique comes back with:
- Fix: the headline describes features, not the outcome; there's no proof above the fold; the page has two competing CTAs.
- Fine: the pricing table and the FAQ.
- Stop worrying about: page length. The current length is fine for the price point.
Those become three constraints for the rewrite. The new draft changes the headline, adds a proof line and drops one CTA, and leaves the pricing table, FAQ and length alone. One focused rewrite replaces several rounds of open-ended polishing.
Critique-first vs other prompting patterns
| Pattern | When the judging happens | Whose standard | Best for |
|---|---|---|---|
| Write, then "make it better" | After, open-ended | The model's default | Quick, low-stakes copy |
| Self-refine | After each draft, by the same model | The model's own | Tasks with clear right answers |
| LLM-as-a-judge evaluation | After, to score or compare outputs | A rubric you supply | Testing prompts or models at scale |
| SDC Loop | At the end of a plan → draft cycle | One expert's, throughout | New deliverables from scratch |
| Critique-First Prompting | Before any writing | An expert's | Improving a brief, an old version or a draft |
How to keep an AI judge honest
The same biases the research found will show up in your critiques. A few habits counter most of them:
- Give criteria, not vibes. "Is this good?" invites the judge's own taste. Three to six named criteria keep it on yours.
- Evidence before verdict. Ask it to quote the passage before it says pass or fix.
- Swap the order when comparing. If you're judging two versions, run it twice with the order reversed. If the winner changes, treat it as a tie.
- Don't reward length. Tell the judge that shorter is fine if it meets the criteria. Verbosity bias is real.
- Separate the judge from the writer. Self-enhancement bias means a model tends to favor its own style. A judge that carries someone else's standard is a better check than the model grading its own draft.
- Keep a human on the final call for anything that goes to customers, regulators or the board.
Running it with Callstand
The definition of Critique-First Prompting asks for an expert Callstand because the critique is only as good as the standard behind it. On Callstand, you build that standard once: you describe your areas, then calibrate by reading realistic test answers and writing what you'd change. Connect it to Claude, ChatGPT, Cursor or any MCP app, and you can ask for three kinds of help: plan it, make it, or check it.
"Check it" is the critique: "Tell me what's wrong first." You paste your work and get what to fix, what's fine, and what to stop worrying about. You don't have to name the call; your AI picks the right one for what you asked. So the critique-first version of a request is just a different order of asking: check the brief or draft first, then ask for the piece with the findings attached. Calls cost $0.50 each from prepaid credits, according to Callstand's How it works page on November 11, 2026.
If you write the same criteria into every chat today, that's the Brief-Once Workflow problem: write them down once and let every tool reuse them.
Frequently asked questions
What is Critique-First Prompting?
Critique-First Prompting is a prompting pattern coined by Callstand: Ask an expert Callstand to judge the brief or the existing draft before asking any model to write, so generation starts from known problems. Instead of asking for a draft and then fixing it, you have AI find what's wrong with the brief or the current version first, and pass that list of problems into the writing prompt as constraints.
What is LLM as a judge?
LLM-as-a-judge means using a large language model to evaluate output, such as scoring an answer, comparing two responses or reviewing a draft against criteria, instead of a human doing it. A 2023 study found GPT-4 as a judge agreed with human preferences over 80% of the time, about the same rate at which humans agree with each other.
Is LLM-as-a-judge reliable?
It's useful but biased. Research has documented position bias (favoring whichever answer comes first), verbosity bias (favoring longer answers) and self-enhancement bias (favoring the judge's own style). You get more reliable judgments with specific criteria, evidence before a score, and swapping the order of anything you compare.
Can AI give good feedback on writing?
Yes, when it has a standard to judge against. Asked to "improve this," a model tends to make cosmetic edits. Asked to check a draft against named criteria, like whether the answer comes first, whether claims are sourced or whether the tone fits the reader, it gives specific, usable feedback.
How is Critique-First Prompting different from the SDC Loop?
The SDC Loop (Strategize → Draft → Critique) runs a full cycle from plan to review. Critique-First Prompting moves the judgment to the very start: when you already have a brief or an existing draft, you critique it before anything new is written. The two combine well, since the critique-first findings feed the strategize and draft steps.
Do I need Callstand to use Critique-First Prompting?
No. You can run a manual version in any AI tool by pasting your criteria and asking for a critique before asking for a draft. The definition calls for an expert Callstand because the critique is only as good as the standard behind it, and a Callstand carries a specific person's standard into every chat.