Session Recorder → Session Recorder with Cursor

Workflow

Debug browser bugs with Cursor

Cursor's agent can read anything in your workspace. Put the recording in the workspace and the agent stops guessing what happened in the browser.

The gap Cursor cannot close on its own

Cursor indexes your codebase. It does not see the request your app sent, the response it got, or the exception the browser threw. So a prompt like “the dashboard shows stale numbers after saving” makes the agent search for plausible causes rather than the actual one — and plausible causes are how you end up with a confident patch to the wrong file.

The workflow

1. Record the bug in Chrome

   Open the Session Recorder side panel, click Record, and reproduce the problem. Narrate while you work — spoken notes are transcribed and land on the timeline at the moment you said them.

2. Export at Compact

   Compact targets roughly 50k tokens: JSON bodies become shape summaries, repeated calls collapse, and screenshots are kept for errors and annotations. It is the level that fits comfortably alongside the rest of Cursor's context.

3. Unzip into the project

   unzip ~/Downloads/stale-dashboard.zip -d .bugs/

   Anything inside the workspace folder is reachable by the agent. Add .bugs/ to .gitignore — recordings can contain real data from your app.

4. Attach it to the Agent

   In the Cursor chat, reference the file directly with @.bugs/stale-dashboard/report.md, or drag the folder into the composer. Then:

   @.bugs/stale-dashboard/report.md
   
   This is a recording of a browser session against this app. [mm:ss] timestamps
   join interactions, network requests, console output and my narration;
   blockquotes are my own notes.
   
   Find the first real failure, trace it into this codebase citing timestamps and
   the specific requests, and propose the smallest fix. Ask before editing.

5. Let it verify

   After the fix, record a second short session of the same flow and attach it. A “before” and “after” recording is the cheapest regression check there is.

Add a rule so you only explain it once

If you record often, put the interpretation in a project rule instead of every prompt. Create .cursor/rules/session-recordings.mdc:

---
description: How to read Session Recorder bug reports
globs: .bugs/**/report.md
---

Files under .bugs/*/report.md are browser session recordings.
- [mm:ss] timestamps are the join key across interactions, network, console,
  screenshots, and the voice transcript.
- "##" headings mark navigations and tab switches.
- "xN similar" means repeated identical requests were collapsed.
- Blockquotes are the user's own markers and notes: highest-priority signal.
- Errors are prefixed with a warning sign and link to the request that caused them.
When a needed body was trimmed, say which [mm:ss] event to re-export instead of
guessing at its contents.

What is in the export

The zip contains report.md (the agent's entry point), session.json (the same events, machine-readable), MANIFEST.md, and the assets: screenshots, full network bodies, audio, and any files you uploaded to the app. Read a complete example.

Local by default

Nothing is uploaded. The session lives in your browser's IndexedDB and leaves only as a zip you save. Secrets are masked at capture time, not after the fact. How the local-first design works.