Prompt Debugging: Fixing a Bad Answer
Diagnose why an AI answer went wrong and apply the right fix instead of rephrasing at random.
A taste of a lesson
I told it 'no more than 100 words' and it gave me 180. I tried three times. What is going on?
Two things are likely. Models do not count words precisely, so limits are approximate, and a long prompt with many instructions can bury the limit. Try this: put the limit as the very last line, make it concrete, such as 'about 5 short sentences, under 100 words', and remove any instruction asking for lots of detail, which pulls the other way. If you need an exact count, check it with a word counter afterwards. Can you paste your prompt so we can look for an instruction that conflicts with the limit?
Written by the teacher as an example. In your lesson the tutor answers your own questions, and like any AI it can be wrong.
What you will be able to do
- Classify a bad answer into a likely failure category
- Apply a matching fix and change one thing at a time
- Recognise when a task is beyond a model's reliable ability
- Keep a simple log that turns fixes into reusable habits
Lesson plan
- 1 Look before you fix Collect the exact prompt, answer and expectation before changing anything. Start
- 2 The failure categories Learn the common ways answers fail and the sign of each. Start
- 3 Fixes for wording problems Apply targeted fixes for format, constraints and missing context. Start
- 4 Facts, refusals and drift Handle the failures that rephrasing cannot fix. Start
- 5 When the tool is the wrong tool Spot tasks a language model cannot do reliably and choose another method. Start
- 6 Your debugging log Turn fixes into habits by recording what worked. Start
Try asking
About this tutor
Everyone gets bad answers. The useful skill is working out why. This tutor gives you a simple diagnostic routine: was the task misunderstood, was context missing, was the format wrong, is it a factual error, did the model refuse or hedge too much, or has a long conversation drifted? Each cause has a different fix, and some failures mean the task is beyond what a language model can do reliably. You practise on real broken answers you bring, change one thing at a time, and keep a short log of what worked. Suitable for any level, from new users frustrated by rambling replies to regular users who hit stubborn problems.
Reviews
4.5
2 ratingsSample
- Fatima Z.Sample
Clear and calm. Good on what models simply cannot do. I wanted a bit more on refusals, which come up in my research work.
- Pedro N.Sample
The category list is on a sticky note on my monitor now. Realised most of my bad answers were drift from huge chats, not bad prompts.
About the teacher
Teaches prompting one technique at a time, with plenty of before and after examples
9 tutors 339 lessons taught Sample
I teach people to write prompts the way I once taught people to write briefs: say what you want, who it is for, and what good looks like. I came to AI from editing and corporate training, so I care more about clear thinking than clever tricks. In my lessons we take one technique at a time, try it on...
See Chiara's profile and tutorsMore like this
Other tutors on the same or nearby topics.