Your Prompt Isn't Working: Debug It With Three Models
A fifteen minute routine that tells you whether to rewrite the prompt or change the model.
TL;DR
Run the failing prompt on two or three models at the same time. If they all miss in the same way, the prompt is the problem, and it is usually missing context, a missing output shape, two jobs in one ask, or no success criteria. If one model gets it right, keep the prompt and change the model.
Is the problem the prompt or the model?
Send the same prompt to two or three models at once and read the failures together. If every model misses in the same way, the prompt is the problem. If one model gets it right and the others do not, the prompt is fine and you have a model fit problem.
That one test removes most of the guesswork. In aiDex, Compare fires a single prompt at several models in parallel and lays the answers side by side, so the pattern shows up on one screen instead of across three browser tabs. The full walkthrough lives in how to compare AI models side by side.
| What you see | What it means | What to change |
|---|---|---|
| Every model misses the same way | Prompt problem | Add context, output shape, criteria |
| Every model answers differently | Ambiguous prompt | Constrain scope and format |
| One model nails it, others do not | Model fit problem | Keep the prompt, switch the model |
| Answers are correct but unusable | Format problem | Ask for a length and a structure |
What actually breaks a prompt?
Four things, roughly in order of how often they show up.
Missing context. The model cannot see your company, your customer, or last week's decision. Anthropic's guidance is to treat the model like a talented new colleague on their first day: it needs the background you take for granted, spelled out (Anthropic prompt engineering docs).
No output shape. "Write about our pricing" has no shape. "Write 120 words, three bullets, aimed at a finance director who has not heard of us" does. Shape is what makes an answer usable instead of merely correct.
Two jobs in one ask. Analysis plus rewriting plus a recommendation in a single prompt usually gets you a weak version of all three. Split them into separate turns.
No success criteria. If you have not said what a good answer looks like, you cannot tell the model it missed, and you cannot score two candidate rewrites against each other.
OpenAI's prompting guide makes a related point about placement: keep role and tone in the system message, and keep the task, the constraints, and the examples in the user message (OpenAI prompt engineering guide).
How do I debug a prompt step by step?
Five steps, about fifteen minutes.
1. Reproduce it in Solo. Run the failing prompt once, unchanged, and write down the specific defect: wrong scope, invented facts, wrong tone, too long. "It was bad" is not a defect you can fix.
2. Split test in Compare. Run the same unchanged prompt across three models, for example GPT-5.4, Claude Opus 4.8 and Gemini 3.1 Pro. Read the three failures next to each other and apply the table above.
3. Score three rewrites in Judge. Write three variants that each fix one suspected cause, then let Judge grade them against your stated criteria and explain the ranking. You learn which change carried the weight, which is the part you keep.
4. Rebuild it in Pipeline. Chain Draft, Critique, Revise and Polish so one model writes the improved prompt, a second attacks it for ambiguity, and a third tightens it. Prompts respond to editing the same way documents do, as covered in building an AI pipeline.
5. Keep the winner in Teams. Park the final prompt in a standing team chat so the next person does not repeat the whole investigation. Which mode to reach for first is mapped out in when to use each aiDex mode.
How do I know the rewrite is genuinely better?
Compare it against the original on the same models in the same session, not against your memory of yesterday. Run old and new back to back, on the same inputs, and check whether the defect you named in step one is gone. If two models now agree where they previously diverged, that convergence is a useful signal, and getting consensus from several AIs explains how far to trust it.
Keep the criteria written down. A rewrite that reads nicer but still ignores your word limit has not fixed anything.
What if the prompt contains confidential data?
Debug it against a local model first. Ollama runs on your own machine, so the failing prompt and any attached documents never leave it, and you can still iterate on structure before you take the finished version anywhere else. Use your own provider keys or the ones we manage, and pick the models you want.
For the wider pattern this fits into, start with multi-model AI workflows.
The aiDex Team · Multi-model AI platform
aiDex is a multi-model AI platform that lets you query several AI models at once, compare their answers, run consensus picks, and chain models in pipelines or open team chats. Use your own provider keys or the ones we manage, and pick the models you want.
Frequently asked questions
How do I know if the problem is my prompt or the AI model?
Run the same prompt on two or three models at once. If they all fail in the same way, the prompt is at fault. If one model answers well and the others do not, the prompt is fine and you have a model fit problem, so keep the wording and switch models.
What is the most common reason a prompt fails?
Missing context. The model cannot see your company, your audience, or the decision behind the request. Adding the background you take for granted, plus a clear output shape, fixes a large share of failing prompts on the first rewrite.
Should I fix everything in my prompt at once?
No. Change one thing at a time and score the variants. If you rewrite context, format and tone together, you learn nothing about which change worked, and you cannot reuse the lesson on the next prompt.
How long should debugging a prompt take?
About fifteen minutes. Reproduce the failure once, run the same prompt across three models, write three targeted rewrites, score them, then keep the winner. Longer than that usually means the task itself needs splitting into smaller asks.
Can I debug prompts that contain confidential information?
Yes. Run them against a local model through Ollama, which keeps the prompt and any attached documents on your own machine. You can work out the structure locally and only move the finished prompt elsewhere if you choose to.
Keep reading
Multi-Model AI Workflows: Why Query All Models at Once (2026 Guide)
One model is one opinion. Here is how to query several at once and get a better answer.
How to Compare AI Models Side by Side
Send one prompt to several models at once, read the answers side by side, and let the output decide instead of the hype.
How to Get a Consensus Answer from Several AIs
Why a synthesized answer from several models beats one model on the questions that matter, and how to get one in two clicks.
How to Build an AI Pipeline: Draft, Critique, Revise
Chain models into Draft, Critique, and Revise stages so each pass improves the last instead of starting over.