record-agent-browser-demo
Skill4 starsChanged 4 months agoMITView on GitHub
Record polished, review-ready web application walkthroughs with Agent Browser, including safe branch/app setup, rehearsal, visible interaction cues, video QA, and optional GitHub PR or issue attachment. Use when a user asks for an Agent Browser video, browser demo, recorded web flow, or video evidence; do not use for general desktop or native-app recording.
In this repository5 other files it gives its agents.
- AGENTS.mdAGENTS.md
- analyze-video-editingSkill
- find-brollSkill
- make-product-videoSkill
- video-editingSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
20 rules. About 0 are common, 15 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈15
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Before running Agent Browser, read references/recording-playbook.md completely and follow its phase order.
- Distinctive: line 13
If the user asks to upload, attach, or post the video to GitHub, also read references/github-upload.md completely before the upload.
- Distinctive: line 17
If the user names Agent Browser, use Agent Browser for both browser interaction and the source recording.
- Distinctive: line 21
Use a fresh named session for the final take.
- Very specific: line 22
Run agent-browser record start --help and ffmpeg -version before the final take.
- Distinctive: line 23
Treat every navigation, modal opening, and major rerender as invalidating element references.
- Distinctive: line 24
Do not expose real passwords, tokens, personal information, or production data.
- Very specific: line 25
Do not upload until the artifact passes visual QA.
- Distinctive: line 32
Read the request, PR/issue description, changed files, existing screenshots, test plan, and repository instructions.
- Distinctive: line 34
Do not record setup work, dependency installation, compilation, authentication, exploratory clicks, or retries.
- Distinctive: line 38
Create an isolated code checkout when needed, install dependencies, start the smallest supported local stack, and wait for the target URL …
- Very specific: line 46
Use a disposable named session to complete the whole shot list.
- Very specific: line 52
Use a new named session and a desktop viewport appropriate to the page.
- Distinctive: line 56
Use a visible DOM cursor or safe key overlay when it materially improves comprehension.
- Very specific: line 60
Stop the recording while FFmpeg remains available on PATH.
- Distinctive: line 62
Run the deterministic helper:
- Distinctive: line 72
Follow references/github-upload.md.
- Distinctive: line 80
Do not report success until all applicable items are true:
- Distinctive: line 87
No sensitive or unrelated content is visible.
- Distinctive: line 89
If uploaded, the remote comment/body and hosted attachment URL were verified.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Record Agent Browser Demo
Produce a concise video that lets a reviewer understand and verify a web change without reconstructing the workflow themselves.
Required reading
- Before running Agent Browser, read references/recording-playbook.md completely and follow its phase order.
- If the user asks to upload, attach, or post the video to GitHub, also read references/github-upload.md completely before the upload.
Non-negotiable invariants
- If the user names Agent Browser, use Agent Browser for both browser interaction and the source recording. Do not silently substitute Playwright, computer-use automation, or a synthetic video.
- Preserve existing work. Inspect Git status before branch operations. If the checkout is dirty or on unrelated work, use an isolated worktree or clone; never reset, clean, or overwrite the user's changes.
- Reproduce the requested code or deployment exactly. Record the named PR/branch/build, not a nearby environment assumed to be equivalent.
- Rehearse the entire flow in a disposable Agent Browser session before the final take. A first attempt is not a final take.
- Use a fresh named session for the final take. Never use Agent Browser's shared unnamed session.
- Run
agent-browser record start --helpandffmpeg -versionbefore the final take. Installed releases can lag the documentation, and missing FFmpeg may not be reported untilrecord stop. - Treat every navigation, modal opening, and major rerender as invalidating element references. Re-snapshot before the next reference-based action.
- Do not expose real passwords, tokens, personal information, or production data. Use documented local fixtures or clearly synthetic demo values.
- Do not upload until the artifact passes visual QA. A successful
record stopalone is insufficient. - Upload or post only when the user explicitly requested it. Verify the resulting remote comment or body after mutation.
Required workflow
1. Establish the review story
Read the request, PR/issue description, changed files, existing screenshots, test plan, and repository instructions. Convert them into a short shot list with one purpose per state. Prefer a beginning, the changed interaction, its important alternate/error state, and the final result.
Do not record setup work, dependency installation, compilation, authentication, exploratory clicks, or retries.
2. Prepare and verify the application
Create an isolated code checkout when needed, install dependencies, start the smallest supported local stack, and wait for the target URL to be healthy. Keep unrelated existing services untouched. Record the exact commit and URL used.
3. Preflight the recording tools
Confirm the installed Agent Browser version, supported recording syntax, browser health, FFmpeg availability, writable output path, and expected GitHub size/type constraints when an upload is requested. Resolve all failures before recording.
4. Rehearse
Use a disposable named session to complete the whole shot list. Capture snapshot -i output after each meaningful UI change, identify stable semantic locators or fresh refs, confirm keyboard behavior, and note any overlays, scrolling, or viewport changes required.
The rehearsal must prove that the intended final state is reachable. If it exposes a product defect, stop and report it instead of disguising the defect in the recording.
5. Record a clean take
Use a new named session and a desktop viewport appropriate to the page; default to 1440x1000 when no product-specific viewport is required. Navigate and stabilize the page before record start. Record only the rehearsed sequence.
Make actions legible: keep the relevant control in view, show menu contents before selecting, use real keyboard events for keyboard behavior, and pause visually on important states. Do not rely on sleep duration alone to control the encoded video's pace; some Agent Browser versions encode only changed frames.
Use a visible DOM cursor or safe key overlay when it materially improves comprehension. Re-inject overlays after navigation. A key overlay must use an explicit allowlist and must never echo arbitrary text or password-field input.
6. Finalize and inspect
Stop the recording while FFmpeg remains available on PATH. Verify that the file exists, is non-empty, decodes successfully, has the intended resolution, and is within the target upload limit.
Run the deterministic helper:
python scripts/video_qa.py path/to/demo.webm --out-dir path/to/qa-framesVisually inspect every extracted frame. At minimum, confirm the start, changed interaction, alternate/error state if present, and final result. Re-record if any important state is missing, clipped, unreadable, too fast, or contaminated by setup/debug UI.
7. Upload when requested
Follow references/github-upload.md. Post a brief factual description of what the video demonstrates, attach the final video with the requested mechanism, and verify the returned comment/body contains the hosted attachment URL.
8. Clean up narrowly
Close the Agent Browser sessions and stop only the temporary servers or containers created for the recording. Preserve the worktree and artifact unless the user asked for their removal. Never tear down unrelated services.
Completion gate
Do not report success until all applicable items are true:
- The recorded code/version and target URL are identified.
- The worktree or checkout is clean of unintended edits.
- The final take came from Agent Browser in a fresh named session.
- The video exists and decodes.
- Representative frames were visually inspected.
- No sensitive or unrelated content is visible.
- The video communicates the requested behavior without explanation from the operator.
- If uploaded, the remote comment/body and hosted attachment URL were verified.
In the final response, lead with the result and provide the remote URL, local artifact path, recorded commit/environment, and any cleanup state. Mention retries or tooling workarounds only when they affect reproducibility.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

