Prompt Engineering·When Things Go Sideways·7 min read·Sep 5, 2026

Fixing Bad Outputs Fast

A symptom-to-fix reference for the most common ways an AI response goes wrong, from ignored formatting to outright refusals.

The prompt is usually the problem

Every prompt engineer gets responses that miss the brief, ignore an instruction, or wander off-topic. The difference between someone who fights the model all afternoon and someone who fixes it in two tries is a habit: diagnose the specific failure before you touch the prompt again.

Symptom → fastest fix

Ignored a format or length instruction

Move that instruction to its own line at the end, in caps or as a bolded "REQUIREMENTS:" block: buried instructions lose to stronger patterns in the response.

Response is generic, could apply to anyone

Add real specifics: your numbers, your constraints, your industry. Generic input produces generic advice every time.

Confident but wrong facts or math

Ask it to show its work and flag uncertainty, or restrict it to only the information you provided.

Answered a different question than you asked

Rewrite the ask so it can only be read one way, and say what you're explicitly not looking for.

Refused or added heavy disclaimers

Reframe away from language that sounds like a request for professional advice, and add the legitimate context (training material, research, education).

Way too long or too short

Give an exact target: a word count or a comparison ("about as long as a tweet") beats "be concise."

Two examples, side by side

✕ Buried instructions

Can you help me understand how our subscription tiers should work, and maybe keep it kind of brief?

Open in ChatGPT ↗

[A five-paragraph essay, because "kind of brief" and "help me understand" don't compete with the model's default toward thoroughness.]

✓ Instructions that stand out

Explain how a 3-tier subscription model (free, pro, team) should differ in features, not just price. REQUIREMENTS: - Exactly 3 short paragraphs, one per tier - End each with the single feature that justifies upgrading - No pricing numbers, just feature differences

Open in ChatGPT ↗

Free: covers a single user with core functionality, enough to get real value, but every feature nudges toward wanting more. The upgrade trigger: hitting the 3-project limit. Pro: unlocks unlimited projects and priority support for one person doing serious work. The upgrade trigger: needing to invite a second person. Team: adds shared workspaces, role permissions, and usage analytics across the whole group. The upgrade trigger: needing to see what teammates are doing, not just your own work.

Two fast diagnostic tests

The minimal-prompt test: strip your prompt down to the bare request. If it still fails, the core ask itself is unclear, not the extra detail. If it suddenly works, your original was over-constrained.
The role-reversal test: paste your own prompt back to the model and ask: "What's ambiguous about this request, and how would you improve it?" It's often better at spotting the gap than you are, because it just tried to fill it.

Keep a short running list of fixes that worked for your recurring tasks. Prompt troubleshooting compounds: the fifth time you fix "too generic" takes ten seconds, not ten minutes.