“Make this better” is easy to type, but it rarely produces the most useful feedback. The assistant has to guess the audience, purpose, tone, and level of revision you want. The result may sound polished while leaving the real writing problem untouched.
Better feedback begins with a clearer contract. Tell the assistant what the text needs to accomplish, who it serves, and which decisions you still control. Then ask it to identify evidence in the draft, rank the problems, and propose revisions without silently replacing your voice.
What direct feedback should contain
Useful writing feedback is not the same as aggressive criticism. It should be clear about the problem while remaining specific enough to act on.
- A concrete observation. Point to a sentence, transition, claim, or structural choice.
- The effect on the reader. Explain why it matters for comprehension, credibility, pacing, or tone.
- A revision priority. Separate major problems from optional polish.
- A practical next move. Suggest a change or question while leaving the final decision with the writer.
“This paragraph is weak” is direct but not useful. “The paragraph introduces three claims before explaining the main one, so a first-time reader may not know which point to retain” gives the writer something testable.
Give the assistant a feedback brief
Before pasting a draft, provide the minimum context needed to judge it fairly: the intended reader, purpose, desired tone, format, length constraint, desired reader outcome, editing stage, and anything that must remain unchanged.
A reusable feedback prompt
Review the draft for a reader who is [audience]. The goal is [purpose]. Focus on [criteria]. Quote the exact passage behind each comment, explain its reader impact, and rank the three most important revisions. Do not rewrite the full piece or invent facts. Preserve [constraints].
Ask for diagnosis before rewriting
A rewritten paragraph can look better without teaching you what changed. Ask the assistant to summarize the apparent argument, identify supporting passages, flag unexpected shifts, rank issues by impact, and propose one revision strategy for each high-priority issue. Approve the direction before requesting a rewrite.
This sequence also makes generic criticism easier to detect. If the assistant cannot point to relevant text, its feedback may come from a template rather than your draft.
Five prompts for common writing tasks
1. Clarify an argument
State the central claim in one sentence. List each supporting claim and quote the sentence that supports it. Flag claims that lack evidence or appear before the reader has enough context. Do not add evidence or citations.
2. Improve structure
Create a one-line outline of the draft as written. Identify sections that repeat a function, arrive out of order, or do not support the purpose. Recommend a revised order using existing material before suggesting new content.
3. Check tone without flattening the voice
Describe the current tone using three precise adjectives and quote evidence for each. Identify any sentence that breaks that tone. Suggest local edits, but preserve the writer’s vocabulary and do not make every sentence sound the same.
4. Make instructions easier to follow
Test these instructions as if you were a first-time user. List prerequisites, actions, decisions, and expected results. Flag missing steps or ambiguous references. Do not claim to have performed the procedure.
5. Prioritize a final revision pass
Give me a three-level revision list: must fix before publication, should improve if time allows, and optional polish. Quote the relevant text and explain the reader impact. Limit the list to the five highest-value changes.
Evaluate the feedback
AI feedback is a second opinion, not an editorial verdict. Check whether it uses evidence from the draft, understands the intended reader, distinguishes facts from preferences, preserves uncertainty, ranks issues, respects the editing scope, and leaves deliberate voice choices intact.
If a suggestion changes meaning, verify it against the source. If it introduces a factual claim, require a source or remove it. If several suggestions make the prose generic, ask what distinctive language was lost.
Compare assistants fairly
Use the same draft and prompt, save complete responses, and score them against the same criteria: intent recognition, specificity, factual discipline, prioritization, voice preservation, and practical usefulness. Record the product, date, prompt, draft, and scoring method. Separate observed results from product claims.
Try the workflow with Pojia AI
Pojia AI is designed to understand legitimate intent and provide direct answers with fewer unnecessary refusals. Judge the feedback by its evidence, precision, and usefulness rather than by whether it simply agrees with you.
Explore legitimate evaluation prompts · Learn how intent-aware boundaries work · Download for Android