Debugging with AI, one hypothesis per branch
Debugging conversations go in circles because the chat forgets the shape of the investigation. You test a theory, it fails, you test another, and the model happily blends them. On a canvas, each hypothesis is a branch, dead ends stay marked dead, and the fix points back at the evidence that produced it.
fork
One branch per suspect: the race condition, the stale cache, the wrong config. Each investigation proceeds clean, without contaminating the others.
reference
Found the smoking gun in branch two? Reference that exact message while writing the fix in branch four. No re-pasting stack traces.
merge
When two partial findings explain the bug together, merge those branches and ask for the unified diagnosis. The post-mortem practically writes itself.
Put the crime scene in the first message
Error, stack trace, expected versus actual, environment. Attach the log file. Every branch you fork later inherits this context automatically.
Branch hypotheses, not questions
A branch is a theory to falsify, not a chat topic. Name the branch after the suspect, chase it to a verdict, and stop. The canvas shows exactly which theories are alive.
Different models, same evidence
Stuck? Fork the trunk and put a different model on the same evidence. Models have different blind spots; the canvas makes the second opinion cheap.
The fix cites its evidence
Write the fix in a merge draft with the confirming branches in scope. When a teammate asks why this change is safe, the answer is the visible path from symptom to evidence to fix.
Common questions
Can I paste code and logs into NodePad?
Yes. Messages take file and image attachments, and attachments travel into branches and merge drafts so every model call sees the same evidence.
Is my code used for training?
No. We never train models on your prompts and responses, and enterprise deployments can run fully inside your own infrastructure.
No credit card to start. Your first workspace is already waiting.