Debug with useful context
Help the AI find the cause by showing what happened and how to reproduce it.
By the end, you can write a report with location, expected and actual behavior, steps, and evidence.
Step 1: Capture the symptom and action
A screenshot shows what you saw, not necessarily why it happened. A good bug report gives the page, viewport, exact steps, expected result, actual result, and recent change. Like giving a mechanic a repeatable route, describe when the problem appears.
Step 2: Find the relevant checkpoint
The browser Console can show browser errors; the Network panel shows and replies. Repeat the action and inspect the made at that moment: address, status, and redacted message. If filtering a large list is slow, compare the same action with small and large data and distinguish network wait from browser work.
Step 3: Send useful, safe context
Give AI a screenshot and redacted errors; remove passwords, tokens, private keys, and personal . A local address is useful to an agent with a local browser connection. Pasting localhost into remote chat does not grant access to your machine.
Step 4: Verify one cause and one small fix
Ask for a hypothesis supported by the evidence, a small proposed edit, and a repeat of the same reproduction steps. For mobile overflow, use the affected width and find the element wider than its container. A general rewrite or clearing saved data is not justified by an unexplained layout symptom.
At /dashboard, 390px width, signing in then opening the second saved card jumps to the top. Expected: card opens without a jump. Actual: page jumps every time. Use the screenshot and redacted Console/Network evidence. Propose one cause and a small patch, then repeat these steps. Do not regenerate the page. Illustrative before / after (behavior, not ready-to-run code): - Dashboard feels broken + Route + viewport + steps + expected/actual + redacted evidence
If you cannot collect a panel capture, say so. The AI should explain how to get the missing evidence instead of inventing a diagnosis.
Picture it, then try it
Follow one request
Track a parcel through its checkpoints. A button starts the trip; it does not prove the parcel arrived.
Browser button → request: bookId + action → server: verify person + permission → database: save permitted result → reply: saved request ID → browser: show confirmed result
The page says Saved, but you cannot find the . Name two things to check before asking AI to recreate the database.
Compare your reasoning with the worked answer
Repeat one . Inspect its and which database/environment it targeted. Then check the read filter and account permissions. A missing in one view could mean a rejected save, the wrong environment, or a view that hides an existing result. Ask AI to report evidence at each checkpoint.
Sources and further reading
The AI cannot reproduce your dashboard jump. What context should you send next?
Choose one answer.