I'm tired of the AI just rewriting my text in a generic way. I want it to actually critique my writing for clarity, tone, and impact. How can I change my prompting style to make it act like a brutal editor rather than a rewrite bot? Do I need to be more aggressive in my instructions? I'm looking for a 'critique' framework that forces the model to explain *why* something is weak before offering a correction.
Effective critique from a large language model requires a prompt structure that explicitly prohibits rewriting in the initial response while requiring a diagnostic analysis of specific text excerpts and a detailed explanation of identified weaknesses.
4 answers
Claude is fundamentally an autoregressive model optimized for helpful, instruction-following behaviors, which defaults to rewriting because that is the most common user expectation. To pivot the interaction toward critique, you need to shift the model's objective function through your prompt architecture.
Consider this a Zero-Rewrite Constraint. You must clearly define the Persona and the Output Schema to override its tendency to rewrite. If you do not explicitly forbid the output of revised text in the first step, the model will naturally gravitate toward it to satisfy its training on helpfulness.
Try structuring your prompt with these explicit directives:
- Define the role: You are a brutal literary editor. Your goal is not to improve the text, but to identify latent flaws in logic, clarity, and tone.
- Prohibit revisions: Do not rewrite any portion of the input in your initial response.
- Require evidence: Quote the specific text being critiqued and provide a granular explanation of why the passage is ineffective.
- Ask for an impact assessment: Describe the likely perception of the reader when encountering these weaknesses.
By defining the interaction as an iterative audit rather than a completion task, you gain the specific feedback you are looking for.
When dealing with LLMs, the root of the problem is often an ambiguous system instruction that defaults to generative output rather than analytical processing. If you want a critique, you must impose a restrictive framework that prevents the model from altering your source text until a diagnostic phase is complete.
I recommend implementing a Protocol-First prompt structure. Treat the prompt like a functional specification for a security audit. By explicitly defining the model's role as an editor who is prohibited from modifying content in the initial response, you force it to adhere to your requirements.
Use the following constraints in your system prompt or initial turn:
- Phase 1: Diagnosis. The model must output a list of identified issues, including clarity gaps, tone inconsistencies, and structural weaknesses, cited against specific line numbers or excerpts from the source text.
- Phase 2: Rationale. For every identified issue, the model must provide an evidence-based explanation of why the current text fails to achieve the desired impact.
- Phase 3: Modification Request. Only after the user approves the analysis should the model generate a suggested rewrite.
By enforcing this procedural separation, you minimize the risk of the model prioritizing creative generation over technical assessment.
Stop asking for a rewrite and stop asking for a critique. If you ask for a critique, you get fluff. If you ask for a rewrite, you get a hallucinated version of your own thoughts. You need to enforce a structured audit.
My team uses a standard template for all internal review processes. We treat the text as an input artifact and the AI as a quality control inspector. It is all about the parameters you set. If the input instructions are vague, the output will be garbage. It is simple Lean logic: tighten the process to improve the quality of the output.
Use this format in your prompt:
- Analyze: Evaluate the text based on the following metrics: clarity, conciseness, and tone.
- Audit: Create a table with three columns: 'Excerpt', 'Identified Issue', and 'Root Cause'.
- Constraint: No rewriting permitted. No suggestions permitted. Only identify and explain.
Once you get the audit, then you can decide if a rewrite is even necessary. Usually, when you see the critique laid out in that table, you will realize that you can fix the issues yourself much faster than trying to prompt your way through a machine generated revision.
Thanks for the advice, Shahistha Sullad. I have been struggling with fluff in my outputs, so using your structured audit template seems like a very direct and helpful way to fix that.
From an NLP research perspective, the issue is that Claude's attention mechanism is heavily weighted toward completion rather than evaluation. If you just ask for a critique, the model is still in 'creative mode.' You have to force it into 'analytical mode' by manipulating the expected output schema.
You need to use schema-based prompting. By forcing the AI to output in a rigid JSON or structured list format, you limit its capacity to default to prose-heavy, generic rewrites. If the model is forced to fill in specific slots like 'Reason for Weakness' or 'Structural Flaw', it has less room to just write a better version of your text.
Add this to your prompt:
- Role definition: You are acting as a data processor for a text quality audit.
- Output schema: You must return your analysis as a set of key-value pairs or a structured table.
- Hard constraint: Any response containing a rewritten version of the text will be considered a failure of the instructions.
By adding this hard constraint, you effectively prune the branch of the model's potential outputs that lead to a rewrite. It is a technical fix for a linguistic problem.
That is a very precise technical explanation, Frank Evans. I must admit, I am slightly overwhelmed by the complexity of schema-based prompting, but I will certainly try to implement your rigorous constraints.
I really appreciate your perspective, Shahistha Sullad. Using a table for root causes is a brilliant idea. I am still learning, but this systematic approach will definitely help me improve my prompting technique.