Screenplay (agent-screenplay) + Claude CodePartly tested🌏 World

Georges Petrov has Claude Code film a demo video for each docs page with open-source Screenplay, and a nightly check goes red when a button is renamed

A short video on every feature page of SaaS docs helps, but each takes an hour to make and goes stale after the next release. Georges Petrov (@georges_petrov) writes each video as a small script: an agent explores the feature in a demo account and writes the steps, and the open-source tool Screenplay clicks through them in a headless browser, films and edits the take (smoothed cursor, zooms that follow clicks, captions, optional voice). The same script is a test: check replays it fast without filming and fails when a button was renamed or removed, so a nightly CI job tells you which video is out of date. On Linux we installed Screenplay 0.1.7, filmed a 22-second 1080p video of its bundled example app, then renamed the "Take over" button; check failed at step 3 with exit code 2.

Visit site Screenplay (agent-screenplay) + Claude Code ↗
Pricing
Free
Region
Global
Where it's from
Overseas
Added
Nov 14, 2026
Tested on
Nov 14, 2026
Version we used
agent-screenplay 0.1.7 (npm), Chrome Headless Shell 159; author used Claude Code with Fable 5.2
Categories
VideoCoding
Original author
Georges Petrov @georges_petrov ↗
Our test
Partly tested
Cost / time
Screenplay is free and open source (MIT), has no AI inside and needs no model key. Writing scripts and reviewing storyboards uses your agent's quota in Claude Code, Codex or Cursor; voice is optional and billed by Cartesia or ElevenLabs.
Source
X @georges_petrov ↗GitHub gmpetrov/agent-screenplay ↗

The task

Give every feature page in your product docs a sub-minute walkthrough video, and find out automatically which videos no longer match the product after a UI change.

Steps

  1. With Node 20+, run npx agent-screenplay init. It downloads Chromium once, adds Screenplay's MCP server to Claude Code (and Codex or Cursor if present), and installs a guide that teaches the agent what makes a good video. The author starts claude --model claude-fable-5-2.
  2. Set up a demo account or staging that looks lived in; never film production (a Send step really sends). Sign in once with npx agent-screenplay login <your app URL> -o videos/.screenplay/auth.json and keep that file out of git.
  3. First have the agent read docs/ and propose one video per page that shows a user action, as a table in videos/shotlist.md, without filming. Cut, merge or split rows yourself. His rule: one video, one idea, under 60 seconds.
  4. Set a house style once in videos/screenplay.config.json (background, padding, corners), or pick it on a frame with npx agent-screenplay style.
  5. Write the rules for a good docs video as a skill (.claude/skills/docs-video/SKILL.md): narrate with the page's own words, say what the viewer gets, check before recording, review the storyboard after each take, three takes at most.
  6. Have the agent start one subagent per shotlist row, four at a time. For voice, enable it in the config and set a Cartesia or ElevenLabs key; without a key, generate audio files yourself and attach one to each step.
  7. Host the MP4s in a bucket behind a CDN and keep only the scripts in the repo. Run npx agent-screenplay check against staging every night; a failure names the page and the step.

Prompt

Read every page in docs/. For each page that shows something a user does in the app, propose one video: the page, the one thing a viewer should be able to do after watching, the starting URL, the 3 to 8 actions that show it, and the result to end on. Skip concept pages and the API reference. Write it as a table in videos/shotlist.md. Don't film anything yet.

What the author says

Georges Petrov @georges_petrov (X article, 12 Nov 2026): the video at the top of their docs' "Needs you" page is 56 seconds, 1080p, with voice-over, a smooth cursor and zooms that follow the clicks; nobody opened a screen recorder, an agent explored the demo workspace, wrote the script and filmed it. With Claude Fable 5.2 one prompt films a whole docs folder. Narration is cached by text, so re-rendering unchanged lines costs nothing, and changing a line only needs render, not a new take. When he renamed "Take over" to "Assign to me" in Screenplay's example app, check failed and saved a screenshot of where it stopped. Screenplay is his own open-source tool (MIT); the article also pitches his Brigade app (free locally, $99/month per workspace in the cloud), which can re-film broken videos and open a PR after each push to main.

Our test

On 14 Nov 2026 on Linux (Node 24.22.1) we ran npm i agent-screenplay (0.1.7), approved ffmpeg-static's install script, and npx agent-screenplay init --no-mcp --no-skill downloaded Chrome Headless Shell 159 (125.7 MiB). With the bundled examples/handoff-demo.json (8 steps against examples/demo-app), check passed all 8 steps in about 3 s; record produced handoff-demo.mp4 (22.3 s, 1920×1080, 30 fps, about 5.2 MB) and a storyboard image in about 33 s total, with visible zooms and captions. We then replaced every "Take over" in demo-app with "Assign to me" and ran check: it failed at step 3 with "nothing visible matches", listed the visible buttons (including Assign to me), saved a screenshot of where it stopped, and exited with code 2, enough to turn CI red. We did not connect Claude Code or Fable 5.2, have an agent write scripts, film real docs or run takes in parallel, and did not test voice or Brigade.

What we found when we tried it:Recording and the staleness check work as the article says: one script produces a 1080p video with zooms and captions, and after a button rename check fails and names the step. How well an agent writes scripts and reviews storyboards needs a run with your own agent.

How to get in

Screenplay's code and README: https://github.com/gmpetrov/agent-screenplay .

Who it suits

Teams with a web app and feature docs who want a video per page; easiest if you already have staging or a demo account.

Pitfalls

  • check and record really click; point them only at a demo account or staging.
  • Parallel takes can change each other's data; give each video its own project and customer, and film the ones that change shared data alone at the end.
  • The saved session expires and needs refreshing; check only catches broken flows, so re-film everything after a redesign.
  • Our npm version doesn't run dependency install scripts by default; we had to run npm install-scripts approve ffmpeg-static and npm rebuild to get ffmpeg.
  • Desktop-app recording needs a real screen; on a headless server only websites work.

More in this category