Using Claude to review your own work can expose blind spots, but only if you separate drafting from judging. A generic request for feedback invites generic encouragement. A useful review gives Claude a defined reader, a concrete goal, evidence to inspect, and criteria that can be tested. The aim is not to let AI decide whether your work is good. It is to generate a second set of observations that you can verify before the work goes public.
A Complete Review Workflow
Start by saving your original draft. Then tell Claude what the work is, who it is for, what action or understanding it should create, and what must not change. Ask for diagnosis before rewriting so you can see the reasoning instead of receiving a polished replacement with hidden tradeoffs.
- Set the brief: State the audience, objective, format, and non-negotiable facts.
- Choose criteria: Review logic, clarity, evidence, tone, and audience fit separately.
- Request citations: Ask Claude to quote each sentence or section that triggered a concern.
- Rank findings: Separate critical problems from optional preferences.
- Revise selectively: Accept only changes you can explain and defend.
- Run a final check: Confirm the revision still sounds like you and preserves every verified fact.
Reusable Review Prompts
For a first pass, use: 'Act as a skeptical editor. Review the draft for the audience and goal below. Do not rewrite it. Return a table with the quoted passage, the problem, why it matters, and the smallest useful change. Mark uncertain comments as assumptions. Audience: [audience]. Goal: [goal]. Criteria: clarity, logic, evidence, tone, and missing objections.'
For bias control, follow with: 'Now argue against your own review. Which comments are matters of taste rather than real defects? Which recommendation could weaken my voice, remove necessary nuance, or introduce an unsupported claim? Keep only findings that survive this challenge.'
For a final proof, ask: 'Compare the revised draft with the original. List any facts, promises, qualifications, names, numbers, or links that changed. Do not assume a change is correct. Flag each one for human verification.'
Common Failure Modes
Claude can sound certain when it lacks context. It may reward conventional structure, flatten an unusual voice, invent a stronger claim, or optimize for persuasion when accuracy matters more. It can also rewrite too early, making it difficult to tell which suggestions were substantive. Avoid prompts such as 'make this better' or 'be brutally honest.' They define an attitude, not a review standard.
- Too agreeable: Require at least three specific concerns and evidence from the draft.
- Too harsh: Ask it to distinguish defects from preferences and explain the consequence of each issue.
- Voice drift: Provide two short examples of your accepted writing and forbid imitation of generic marketing language.
- Fact drift: Tell Claude not to add facts and require a change log for names, numbers, claims, and quotations.
A Worked Before-and-After Example
Suppose a project update says: 'The launch went well and users loved the new onboarding. We should invest more in it next quarter.' A weak review might call the sentence clear and suggest making it more exciting. A criteria-based review identifies two unsupported claims: 'went well' has no success measure, and 'users loved' has no source. It also spots a decision leap from positive sentiment to more investment.
A stronger revision is: 'In the first two weeks, onboarding completion rose from 58% to 71%, while five of twelve interviewed users described setup as easier. Before increasing next-quarter investment, we should confirm whether completion remains above 70% and whether activated users retain at the same rate.' The revision is better because the claims became measurable, the evidence became visible, and the recommendation gained a test.
Keep the Final Judgment Human
Review Claude's comments one by one. Ask, 'What evidence supports this? Does the change serve the stated reader and goal? What would be lost if I accepted it?' If you cannot answer those questions, leave the original text in place or gather more information. A dedicated Claude Project with your style guide, approved examples, and factual references can improve consistency, but it does not remove the need to verify the result.
The most reliable process uses Claude as a structured dissenting reader, not an authority. Preserve the original, demand evidence, challenge the critique, and approve changes deliberately. That workflow reduces both your own bias and the AI's tendency to replace judgment with confident prose.