Ask the AI Reader about one specific behavior
How to phrase questions about values, async bugs, and patches, which context to add, and how to check the answer.
“Explain this project” gets you an overview. It rarely tells you why a particular branch ran or whether a change is safe. You’ll get more out of the AI Reader if you ask about one behavior and say what you want checked.
Say what happened and what you expected
A good first question says what you did, what happened, and what you expected instead. Add the file or selection if you know it. If you don’t, describe the message or action you saw and ask the AI Reader to find where it starts.
For example:
This report returns four rows with
REPORT_LIMIT=4, even though the config file says six. Trace the limit from input to output. Show which source wins and what happens when the environment value is empty.
This question has a clear end: which source is picked, where it’s validated, and where it’s used. It also points to a small test you could write. The configuration walkthrough shows what a complete answer covers.
Choose a starting context
| Context | Good for | Still check yourself |
|---|---|---|
| Selection | Explaining one expression or branch | What the caller assumes, surrounding conditions |
| File | Behavior spread across several methods | Imports and outside state |
| Folder | You know the subsystem but not the entry point | Which files actually take part |
| Git comparison | Reviewing specific commits | Base and head direction, older context |
What you add is only a starting point. The AI Reader can open other files in the project you authorized to follow dependencies, so a request may include more than the file you added. If that matters for confidential code, read what gets sent first.
Example questions
For an async UI bug:
Selecting A and then B sometimes leaves A’s contents on screen. Trace both requests across each await. List the writes in success, catch, and finally, and show what stops an old request from changing the current view.
The JavaScript async race condition tutorial reproduces this exact sequence and uses a request counter to protect the text, error, and loading state. Download its example to run the included tests.
For a patch review:
Review this comparison for changes to cache invalidation. For each possible regression, give the changed location, an order of events that can actually happen, what the user would see, and a test that tells the old and new behavior apart. List existing issues separately.
For an architecture question:
Which component owns this cache, who can invalidate it, and which callers share an entry? Cite the constructors and the code that changes it. Mark anything that depends on runtime configuration.
Each question names what a good answer has to include: an order of events, a test that tells two versions apart, or the code that creates and changes the cache.
Check one claim before going further
Open a key reference in Code Quick Look, and the whole file if you need more. Check whether the cited condition really holds for your input, and whether another caller, override, or error path changes the answer.
When the answer leaves a gap, ask a narrow follow-up:
You found two implementations. Where does this request pick one? If the code alone can’t tell, which runtime value do we need?
The AI Reader can’t run your project’s tests, so a test it suggests hasn’t been run yet. Run it yourself and paste the result back into the conversation.
Follow-ups and side questions
If the next question digs deeper into the same trace, keep asking in the same conversation. To explore a separate direction, use Fork into New Chat on a finished answer. That starts an independent conversation with the context up to that answer and leaves the original thread as it was.
The AI Reader only changes files after you turn on Allow Edits for a run, and later runs and forked conversations start read-only again. Git comparison conversations are always read-only. After an edit, read the diff and run the checks that cover the change.