Prompt Troubleshooting Guide
About this resource
This is the reference version of prompt troubleshooting, organized as a lookup table you scan rather than a path you follow. Where the Prompt Improvement Flowchart walks you through a single problem by branching, this lists the recurring problems side by side with their causes and fixes, which suits somebody who has met most of them before and wants to find the right one quickly.
It covers the persistent problems rather than the one-off ones: output that drifts over a long conversation, instructions that are followed once and then forgotten, formatting that degrades as a response gets longer, the model becoming agreeable rather than accurate, answers that hedge everything, and the specific problem of a model confidently completing a task it should have refused. Each entry names the underlying cause, which is what makes the fix stick.
Built for the problems that keep returning rather than the one-off puzzle. Those are the ones with a structural cause and a durable fix.
How to use it
Keep it alongside the flowchart rather than instead of it. The flowchart is for the moment a single answer disappoints; this is for the pattern you have noticed across a week of work. The most valuable entry for most people is conversation drift, because it affects sustained work and almost nobody realizes it is happening.
Inside the download
Why long threads degrade and how to reset without starting over.
Why a constraint is followed once then abandoned, and what to do about it.
Why structure degrades as responses get longer.
When the model starts agreeing rather than assessing, and how to break it.
Common questions
Because the earliest content stops fitting in the context window, so the instructions and material you established at the start gradually stop influencing responses. The fix is to restate the essential constraints periodically, or to start a fresh conversation carrying a short summary forward.
The instruction is competing with everything that came after it. Constraints given once at the start weaken as a conversation grows, so anything that genuinely matters belongs in custom instructions or restated near the request rather than assumed to persist.
Because agreement is a strong default in how these systems were trained, and it intensifies once a conversation has established a direction. Ask explicitly for the case against, or present the position as somebody else’s rather than your own, which removes the pull toward agreeing with you.
Free. No email list, no follow-up sequence.
More from the resource library, chosen because they pair well with this one.
A decision tree for fixing a disappointing answer: diagnose the symptom, then change one thing.
Ready-made instructions for getting tables, structured summaries, and clean paste-ready output back.
Tokens, context windows, training cutoffs, and why these systems produce confident wrong answers. One page, kept beside a monitor.
Tell us a little about yourself and this document is yours. We don't have an email list. We collect this so we know who's finding our resources useful, and that's it.
You're downloading
this document
Your information is kept confidential and never shared with anyone.
