Don't Ask AI to Write Directly: Use Layered Critique Prompts to Improve Results

Asking AI to write from scratch is quick, but the first result usually still has issues: the opening is too general, arguments are repetitive, examples are lacking in concreteness, or the tone does not match the audience. The problem is not always with the model's capabilities, but rather because we ask for the final result without giving it a chance to test the draft.

One more practical approach is to use layered critique prompts. The flow is simple: ask AI to create a draft, then ask it to identify weaknesses in that draft, and finally request improvements based on the identified critiques. This technique does not guarantee that all facts will be correct, but it helps us view the work from multiple angles.

Why is one prompt often not enough?

Prompts like “Write an article about the benefits of AI for small businesses” provide a goal that is too broad. AI has to guess who the audience is, how long the article should be, what perspective to take, relevant examples, and the desired writing standards.

When all those decisions are handed over at once, the result tends to be safe and generic. AI may construct neat sentences, but it may not precisely address the audience's needs.

Layered prompts break the work into several stages. Each stage has a narrower goal so that we can evaluate the results before moving on to the next stage. In prompt engineering practice, providing clear instructions, output formats, and relevant examples indeed helps guide the model's results more consistently.

Basic structure of layered critique prompts

You can use the following four layers:

1. Production: create an initial draft based on context and objectives.
2. Diagnosis: identify weak, confusing, repetitive, or unsupported parts.
3. Prioritization: select the issues that most impact the writing's goals.
4. Revision: improve the draft without changing what already works well.

The key is not to mix all instructions into one long command. You can run it in one conversation so that AI still has the context of the draft, or separate it into several conversations if you want to assess the results more independently.

Step 1: Request a draft with clear limitations

Start by explaining the desired outcome, not just the topic. Include the audience, purpose, tone, length, and things to avoid.

You are a content editor for small business owners.

Write a draft article of about 800 words on how to use AI to summarize meetings.
Audience: business owners who are not yet familiar with using AI.
Purpose: the audience understands a simple workflow they can try today.
Tone: practical, warm, not exaggerated.
Must include: before and after examples, risks of mis-summarizing, and verification steps.
Avoid: technical jargon, claims that AI is always accurate, and cliché openings.

Create a draft with easily scannable subheadings.

Note that the prompt does not just say “write an article.” It provides context, limitations, and success criteria. If format is important, state it explicitly. For recurring tasks, you can even save this prompt as a template with replaceable parts.

Step 2: Ask AI to be a critic, not a rewriter

After the draft is complete, do not immediately ask for a better version. First, ask AI to identify the problems. This instruction is important so that the critique does not turn into a new piece of writing that makes us lose the rationale behind the changes.

Review the draft above as a critical editor.
Do not rewrite the article yet.

Create a table with columns:
1. problematic section or quote,
2. type of problem,
3. why that problem is important,
4. suggested improvements.

Specifically check:
- whether beginner readers can follow the steps;
- whether the examples are concrete enough;
- whether there are overly broad claims;
- whether there are repetitive sections;
- whether the risks and limitations of AI are explained;
- whether the opening directly addresses the audience's needs.

Requesting categories of critique like this makes the evaluation more focused. AI does not just say “this writing can be improved,” but must show the location of the problems and the reasons.

Step 3: Ask for prioritized critiques

Not all feedback needs to be implemented. Changing one awkward word is not as important as fixing a step that the reader cannot actually perform.

Use a follow-up prompt like this:

From the previous critique, select a maximum of five most important issues.
Rank them based on their impact on the clarity and usefulness of the article.

For each issue, explain:
- the risk if the issue is left unaddressed;
- the minimum changes needed;
- whether that change requires additional information from me.

Do not suggest changes in writing style if the substantive issues have not been resolved.

This prioritization stage helps prevent excessive revisions. In editorial work, a piece that has been changed more does not necessarily become a better piece. Focus on changes that bring the draft closer to the original goals.

Step 4: Request revisions with ground rules

Only after the critiques are understood, ask AI to improve the draft. Provide rules so that the revisions do not change the entire piece without reason.

Revise the draft based on the five prioritized issues above.

Revision rules:
- maintain the purpose, audience, and tone of the writing;
- do not add statistics or product names without sources;
- retain sections that are already clear and concrete;
- mark sections that still need verification with [NEED CHECK];
- after the article, create a list of major changes in a maximum of five points;
- if there is insufficient information, state the gaps rather than guessing.

Instructions like “do not guess” do not automatically free AI from errors. However, such instructions signal that uncertainty should be displayed, not covered up with seemingly confident statements.

When is this technique most useful?

• Writing articles or newsletters: use critique layers to check flow, openings, examples, and claims.
• Drafting proposals: ask AI to identify unproven assumptions and questions clients might ask.
• Creating technical documentation: ask AI to check whether readers have all prerequisites before following steps.
• Preparing presentations: ask AI to test whether each slide has one main message and whether the order makes sense.
• Selecting ideas: ask AI to critique the weaknesses of each option before helping to make recommendations.

Do not make AI the final fact-checker

Critiques from AI remain input, not final judgments. The model can generate critiques that sound reasonable but are incorrect, missing important context, or assuming certain claims are true just because they are stated confidently.

For numbers, rules, product specifications, medical information, legal matters, or facts that can change, verify with primary or trusted sources. If you use AI to summarize documents, cross-check important parts with the original documents.

Think of AI as a fast second editor, not as the decision-maker. It helps identify weaknesses; you remain responsible for determining what is true and worthy of publication.

What you can do now

Choose one task that you usually send to AI. Do not immediately ask for the final result. Run three simple rounds: request a draft, ask for the five most important weaknesses, then request revisions with clear rules.

After that, compare the results with the old way. Is the opening more direct? Are the examples easier to implement? Are there claims that are now more cautious? If the answer is yes, save that prompt pattern and adjust it for the next task.

The biggest improvements often do not come from longer prompts, but from a process that allows AI to generate, review, and then improve its work step by step.

Sources & further reading

• Prompting best practices - Claude Platform Docs
• Prompt design strategies - Gemini API
• Structured outputs - Gemini API

– Rio Yotto @rioyotto