How to Improve a Bad Prompt: A Diagnostic Workflow

A weak prompt is not simply a short prompt. It is a prompt whose requirements are hard to identify, whose context is missing or irrelevant, or whose output cannot be judged against clear conditions. The goal of revision is not to make every prompt longer. It is to remove uncertainty that matters for the task.

Quick diagnosis#

A prompt probably needs work when the response is generic, aimed at the wrong reader, formatted for the wrong use, based on invented assumptions, or inconsistent across repeated attempts. Before rewriting everything, identify which structural element is missing.

Symptoms of a weak prompt#

• The request names a topic but not an action.
• The model must guess the audience, purpose, source facts, or success criteria.
• The requested tone conflicts with the intended situation.
• The output format is described with vague words such as “good,” “professional,” or “detailed.”
• Important limits appear late or contradict earlier instructions.
• The prompt asks for factual claims without supplying sources or a verification step.

The eight-part diagnosis#

Use the same public criteria as the rule-based Prompt Checker. They are structural signals, not a prediction of response quality.

1. Clear goal: What observable action should the model perform?
2. Context: Which facts, source material, and boundaries are necessary?
3. Audience: Who will use or read the result?
4. Role: Would a perspective or professional standard help?
5. Output format: What shape should the result take?
6. Constraints: What measurable limits or exclusions apply?
7. Examples: Would one or more examples reduce ambiguity?
8. Quality check: How should the result verify its own compliance?

A prompt does not need all eight elements every time. A two-line prompt can be sufficient for a low-risk, familiar task. Add only the elements that reduce a real source of error.

Eight practical fixes#

Replace the topic with an action#

Change “customer retention” to “compare three retention experiments and recommend which one to test first.”

Add decision-relevant context#

Include verified facts, available resources, current constraints, and what has already been tried. Do not include confidential material that the selected AI service should not receive.

Name the reader#

“Explain this” becomes more useful when the prompt says whether the reader is a new employee, a technical reviewer, a customer, or an executive.

Use a role as a lens, not a credential#

“Act as a senior lawyer” cannot create legal authority. A safer instruction is “review the text for unclear obligations and list questions for qualified counsel.”

Define a usable format#

Specify sections, order, fields, or a table layout. The format should match the next step in the workflow.

Make constraints testable#

Prefer “120–160 words, one CTA, no unsupported performance claims” over “keep it concise and convincing.”

Add examples only when they teach a pattern#

A relevant example can clarify style or classification. An irrelevant example can anchor the model to the wrong assumptions.

Ask for a compliance check#

Request a short checklist that confirms required sections and labels uncertainty. This does not verify factual truth; it only makes omissions easier to see.

Before and after: rewriting task#

Weak prompt

Rewrite this article to make it better.

Improved prompt

Rewrite the article below for small-business owners who understand the topic but not the technical terminology. Preserve every supported fact and quotation. Improve the order, headings, and transitions without adding new claims. Return: revised headline, 80-word summary, revised article, and a change log listing any sentence whose meaning was materially altered. If a statement cannot be verified from the supplied article, mark it [CHECK].

[ARTICLE]
...
[/ARTICLE]

The revised prompt identifies the audience, source boundary, allowed changes, output structure, and uncertainty treatment.

Before and after: debugging task#

Weak prompt

Fix this code.

Improved prompt

Diagnose the TypeScript error in the enclosed function. The project uses strict mode and Node 22. First identify the smallest likely cause, then propose a minimal patch. Do not change the public function signature. Return: diagnosis, patched code, one regression test, and assumptions that still need confirmation. Do not claim the code was executed.

The improvement does not guarantee a correct fix. It makes the requested reasoning boundaries and evidence visible.

Guided Fix walkthrough#

1. Paste the original prompt into the Prompt Checker.
2. Review “Detected” and “Missing” items rather than focusing only on the score.
3. Open Guided Fix and answer only the questions that affect the task.
4. Compare the improved prompt against the original intent.
5. Remove any detail that the model does not need.
6. Run the prompt in your chosen AI and inspect the actual response.

The Guided Fix is rule-based. It combines your answers with the original text; it does not understand whether your facts are true or whether a requirement is legally or professionally appropriate.

When not to add more detail#

Do not expand a prompt merely to maximize a structure score. Extra instructions can conflict, hide the central task, or make a simple request harder to follow. Keep a prompt compact when the task is low risk, the context is already shared in the conversation, and the output can be checked quickly.

Model variability#

The same prompt can produce different results across models, product settings, dates, and conversation histories. Treat a prompt as a testable interface, not a permanent guarantee. Record the model, date, relevant settings, and observed failure when consistency matters.

Verification checklist#

• Does the revised prompt preserve the actual goal?
• Are all supplied facts safe and relevant?
• Can a reviewer check the requested format and limits?
• Does the prompt distinguish source facts from assumptions?
• Are high-risk claims reserved for qualified review?
• Have you tested the prompt with realistic input rather than a perfect example?

FAQ#