Learn how to improve an AI-generated response

How to Improve an AI Response

The Beginner-Friendly RERUN Workflow

An AI response does not have to be completely wrong to be disappointing.

It might be accurate but generic. The introduction may be too long, the tone may not suit your audience, or the recommended steps may lack practical detail.

Many people respond by starting a new conversation or entering another broad instruction such as:

“Make it better.”

The AI system cannot reliably determine what “better” means to you. It may rewrite useful sections, introduce unsupported information, or change the tone without fixing the actual problem.

A better approach is to inspect the answer and provide controlled feedback.

THE RERUN WORKFLOW

REVIEW → EXPLAIN → REQUEST → USE CONSTRAINTS → NOTE

R — Review the Response Before Requesting Changes

Read the complete response before submitting another prompt.

Evaluate five areas:

Review Area	What to Check
Accuracy	Are important claims, names, numbers, and dates correct?
Relevance	Does the response actually answer the question you asked?
Structure	Is the information organized appropriately for the task?
Tone	Does the wording suit the intended audience?
Safety	Does the task require privacy, security, or qualified human review?

Then separate the response into three categories:

1. Keep: Material that works and should remain
2. Fix: Material that needs revision
3. Verify: Material that must be checked or removed

This prevents a useful draft from being discarded because one section is weak.

For time-sensitive or consequential information: Open reliable sources yourself. An AI system should not be treated as the final approver of the accuracy of its own answer.

E — Explain the Gap Specifically

Feedback should describe the problem—not merely express dissatisfaction.

Feedback Type	Example
Vague	“Make this more useful.”
Actionable	“Keep the five recommendations, but shorten the introduction to two sentences. Add one beginner-friendly example beneath each recommendation. Do not add statistics or product claims.”

The second instruction identifies what to preserve, what to change, and what the system must avoid.

Useful feedback may identify:

• Missing steps
• An inappropriate reading level
• Unsupported claims
• Unclear organization
• Excessive length
• Repeated ideas
• An unsuitable tone
• Missing limitations
• A lack of examples
• Information requiring verification

R — Request One Clear Revision

When possible, revise one meaningful target at a time.

Examples of Focused Revision Requests

• Rewrite the introduction for beginners.
• Convert the explanation into a five-step checklist.
• Remove unsupported claims.
• Add two examples based only on the supplied source.
• Make the tone professional but conversational.
• Reduce the draft to 300 words without removing safety guidance.

Requesting many conflicting changes simultaneously can make the outcome difficult to evaluate.

If the draft has serious factual or structural problems, a larger rewrite may be necessary. Otherwise, targeted revisions make it easier to determine whether your feedback actually improved the response.

U — Use Explicit Constraints

Constraints establish boundaries for the revision.

Constraint	What to Specify
Length	Word count or approximate length
Audience	Who should understand or use the response
Format	Paragraphs, checklist, table, email, report, or other structure
Tone	Professional, conversational, beginner-friendly, concise, etc.
Must Keep	Facts, sections, warnings, sources, or conclusions that must remain
Must Not Add	Unsupported statistics, quotations, prices, features, or assumptions
Review Requirements	Claims requiring sources, verification, or qualified human review

Example

“Revise this for small-business owners who are new to AI. Use a short introduction, five numbered steps, and a concluding checklist. Keep the existing privacy warning. Do not add statistics, prices, or product features.”

Constraints cannot guarantee a correct answer, but they make the requested outcome clearer and easier to assess.

N — Note What Worked

Save successful instructions instead of reconstructing them later.

Your note might contain:

• The type of task
• The feedback that improved it
• Required tone
• Preferred structure
• Necessary safety language
• Frequent problems to check
• A reusable prompt

Over time, these notes can become a personal prompt library.

Privacy reminder: Do not save confidential material in an unapproved system. Keep reusable instructions separate from sensitive customer, employee, or company information.

REVIEW WHAT YOU GOT → EXPLAIN THE GAP → REQUEST THE CHANGE
→ SET THE BOUNDARIES → SAVE WHAT WORKED

Copy-and-Paste RERUN Prompt

Review the previous response against my original request. Keep the parts that already answer the request accurately. Revise only the gaps I list below. Follow the audience, format, length, and tone constraints exactly. Do not invent facts, statistics, quotations, sources, prices, or product features. Mark anything requiring verification as: “Needs verification.” After revising, provide a short change log explaining what you preserved, changed, and could not verify. KEEP: [LIST WHAT WORKS] FIX: [LIST SPECIFIC PROBLEMS] AUDIENCE: [DESCRIBE AUDIENCE] FORMAT: [DESIRED STRUCTURE] LENGTH: [TARGET LENGTH] REQUIRED INFORMATION: [MUST REMAIN] DO NOT ADD: [PROHIBITED ADDITIONS]

Example: Weak Feedback vs. Useful Feedback

Imagine that AI produces a long, generic explanation of how businesses can use automation.

Revision Request	Example
Weak	“Make it shorter and more interesting.”
Useful	“Keep the explanation of human approval. Reduce the introduction to 50 words. Organize the remaining material into four numbered steps for small-business owners. Add one nontechnical example based only on the existing draft. Remove repeated benefits and do not add performance statistics.”

The stronger version defines both the destination and the guardrails.

Five AI Revision Mistakes to Avoid

Mistake	Better Approach
Revising without checking accuracy	Verify consequential claims independently before focusing only on wording.
Replacing everything	Identify what should remain so useful material is not unnecessarily rewritten.
Using conflicting constraints	Make sure requirements such as depth and word count can realistically coexist.
Allowing new claims	Check the revised version again because revisions can introduce new factual claims.
Skipping final human review	Do not treat improved wording as proof of accuracy or suitability.

Keep Consequential Decisions Human

AI can help organize, rewrite, and explain. People should retain responsibility for final decisions involving employment, legal obligations, finances, medical matters, security, or safety.

Remove sensitive information before sharing a draft with an AI system. Follow your organization's policies and use approved tools.

BETTER WORDING DOES NOT AUTOMATICALLY MEAN
BETTER EVIDENCE.

Conclusion

A disappointing first response does not always require a completely new prompt.

Use RERUN:

Review → Explain → Request → Use Constraints → Note

Identify what works, describe the gap, request a focused change, provide clear boundaries, and save the instructions that produce reliable results.

You do not always need a better first prompt. Sometimes you need better feedback on the first answer.