"The checkout button is broken, can you fix it?"
Type that into an AI coding assistant and you'll get an answer. It'll probably be confident, and it'll probably change something. Whether it fixes your actual bug is a coin flip, because the assistant has to guess what "broken" means, where the button is, and what should happen instead.
AI assistants are good at fixing bugs they understand. Your job is to make the bug understandable. Here's how.
The template
Every good bug prompt answers the same questions. Copy this and fill it in:
## Bug
<One sentence: what's wrong.>
## Where
Page: <full URL>
Element: <what was clicked or looked at, and its text>
## Environment
Browser: <name and version>
OS: <name and version>
Screen: <width x height>
## Steps
1. <first thing>
2. <second thing>
3. <what happened>
## Expected
<what should have happened>
## Console
<any errors, copied exactly>
## Ask
Find the cause before changing code. Explain it in a sentence, then propose the smallest fix.
Why each part matters
Bug. One sentence keeps the assistant focused on a single problem. If you have three bugs, write three prompts.
Where. The URL tells the assistant which route or page file to look at. The element text is often the fastest way for it to find the component in your codebase.
Environment. Many bugs only exist at certain screen sizes or in certain browsers. Without this, the assistant can't tell a layout bug from a logic bug.
Steps. Steps turn a symptom into a reproduction. "It broke" is hard to reason about. "It broke after changing the quantity to 3" points straight at the quantity handler.
Expected. This is the part people skip, and it matters most. Without it, the assistant may "fix" behavior that was correct.
Console. An error message with a file name and line number is often the whole answer. Copy it exactly. Don't paraphrase.
Ask. Asking for the cause first keeps the assistant from patching symptoms. It also gives you something to check: if the explanation doesn't make sense, the fix won't either.
Don't write it by hand
Filling in this template from memory is slow and error-prone, and the person who saw the bug often isn't the person fixing it. The good news is most of it can be captured automatically at the moment the bug is reported.
In FeedbackDots, every report has a Copy for LLM button. It puts the report's title, comments, dots and the elements they're on, device details, and console logs on your clipboard as Markdown, structured much like the template above. Paste it into your assistant, add the "Ask" line, and you're done.
Check the fix where it broke
When the assistant proposes a fix, test it under the same conditions as the report: same page, same screen size, same steps. A fix that works on your desktop may not fix a bug that only happens on a phone.
If the fix works, close the loop with whoever reported it. If it doesn't, give the assistant the new behavior and ask it to try again. A second, specific prompt beats starting over.
The short version
Give your AI assistant what you'd give a new teammate: where it is, what it's running on, how to reproduce it, what should happen, and what the console said. The more of that you capture automatically, the less time you spend writing prompts and the more time you spend shipping fixes.
Next Wednesday: what you miss when you only test in a desktop browser's phone mode.