Community post
Make a Bad Prompt Better in Three Edits
By Liam · Published · 2 min read
Editorial standards and correctionsA before-and-after prompting method that changes task, context, and test criteria one at a time.
When a prompt fails, I do not replace it with a giant template. I change the part most likely to fix the observed problem, test again, and keep a record. Here is a three-edit example.
A revision only matters if it beats the first version; Compare Two Prompt Versions Without Fooling Yourself shows a fair way to compare them.
Version 0: vague
Help me make a game tutorial.
Possible problem: the model could produce a long explanation instead of a playable tutorial. I have not defined the player action or what success looks like.
Edit 1: define the task
Draft a 90-second tutorial for a two-player puzzle game. The player must learn to pass one object to a partner.
Now the job has a time limit and a mechanic. I can reject a tutorial that never teaches the pass action. I still may get too much exposition.
Edit 2: add relevant context
Draft a 90-second tutorial for a two-player puzzle game. The players begin in separate rooms and can pass one object through a slot. They cannot speak through the game. Teach the pass action through play, not a text screen.
The separate rooms and no-chat constraint change the design. The model should propose an interaction players can discover.
Edit 3: specify the check
Return three tutorial beats. For each, list what the player sees, what they do, and what confirms success. The last beat must require both players to cooperate. Do not add a new mechanic.
Now I can inspect whether each beat is playable and whether the last one actually requires cooperation. If it fails, I point to the failed beat rather than saying “do better.”
Keep a prompt log
Save each version, the output, and one sentence about why you changed it. Do not compare outputs from different tasks or silently change the success criterion. A controlled sequence shows what helped. This echoes the iterative approach in Google’s prompt guidance.
Try it: Start with a prompt that disappointed you. Make three revisions: action, relevant context, and test. Write the observed failure beside each version. Stop when the result passes your criterion; do not keep polishing for its own sake.
Keep learning: The Prompt Failed. What Should You Change First?, How to Test a Prompt on More Than One Example, and Compare Two Prompt Versions Without Fooling Yourself.
About the author
Liam
Coder and gamer. I test prompts, share what works, and show how to improve AI results for code and creative projects.
View Liam's public profileComments (0)
Loading comments…
Keep exploring
Related from Prompt Engineering Lab
How to Prompt Across Models Without Pretending They Work the Same
Keep the task and success criteria stable while checking each AI provider’s current model-specific guidance.
Compare Two Prompt Versions Without Fooling Yourself
Hold the task and test set constant, change one instruction, and record tradeoffs when comparing AI prompts.
Prompt AI for a Game Idea Without Getting Generic Tropes
A game-design prompt that forces concrete player choices, failure states, and a quick playable test instead of generic story ideas.