Skip to content
DocsThe DemoTake skill
Docs / Agents / The DemoTake skill

The DemoTake skill

One SKILL.md file teaches your agent how DemoTake works, from planning scenes to exporting a checked video. Install it once and your agent stops guessing which tool comes next.

1 minuteClaude Code and any agent

Connecting DemoTake gives your agent the tools. The skill gives it the playbook: which tool to call first, when a scene should end, how to retry just one scene, and how to export something you can ship. Agents that load it make fewer wrong turns and ask you fewer questions.

The skill is a single Markdown file at https://demotake.my/skills/demotake/SKILL.md. The Agents page has it too, with Copy and Download buttons.

Install it in Claude Code

Claude Code loads skills from a skills folder, either for you or for one project:

Scope Path
Every project on your Mac ~/.claude/skills/demotake/SKILL.md
One project (commit it for your team) .claude/skills/demotake/SKILL.md in the project

Download it straight into place:

Terminal
$ mkdir -p ~/.claude/skills/demotake
$ curl -fsSL -o ~/.claude/skills/demotake/SKILL.md https://demotake.my/skills/demotake/SKILL.md
$ ls ~/.claude/skills/demotake
SKILL.md
For a project-level install, use .claude/skills/demotake inside the project instead.
Terminal
mkdir -p ~/.claude/skills/demotake
curl -fsSL -o ~/.claude/skills/demotake/SKILL.md https://demotake.my/skills/demotake/SKILL.md

Start a new Claude Code session afterwards so it loads the skill.

Install it in other agents

  • Agents with skills folders: put SKILL.md wherever your agent loads skills from, in a folder named demotake.
  • Agents with instruction files: paste the contents into AGENTS.md, your editor's rules file, or the project instructions your agent reads at the start of a session.
  • Any language model: point it at /llms.txt, a plain-text map of every docs page, tool and API endpoint, or at /llms-full.txt, all of the docs in one file for agents that prefer to load everything up front.

The skill doesn't replace the MCP connection. Set that up first in Connect your agent.

What the skill teaches

The heart of the skill is the recording loop. Every call names its handles (take_…, scn_…), so the agent always knows which take and scene it's working on.

  1. permissions_status
  2. targets_list
  3. take_createwith scenes
  4. capture_start
  5. take_start_sceneper scene
  6. act
  7. assert_*
  8. take_checkpoint
  9. capture_stop
  10. edit_suggest_cuts
  11. edit_apply
  12. export_renderwait: true
  13. export_validate
  14. share
The order the skill teaches. The scene steps repeat for every scene.

In words:

  1. Check permissions and find the app. permissions_status first, then targets_list to pick the window (or a booted iOS Simulator device).
  2. Plan scenes. take_create with a goal and three to six scenes, each a story beat that ends at a stable screen: "Sign in", "Create project", "Export".
  3. Record scene by scene. capture_start, then for each scene take_start_scene, do the steps, prove the result with assert_text or assert_element, and end with take_checkpoint describing the stable state. If a scene goes wrong, take_retry_scene re-records that scene only.
  4. Tighten the cut. capture_stop, then edit_suggest_cuts to review long still stretches (consent and result screens are protected), and apply the reviewed plan with edit_apply. Every edit returns an id that edit_revert can undo.
  5. Export and verify. export_render with wait: true renders the formats you asked for; export_validate confirms the file opens, has the right length and size, and decodes.
  6. Share the result, or hand you the file paths.

It also covers the "demo as code" loop: take_to_spec saves a finished take as demotake.json, and take_replay with only_changed: true re-records only the scenes whose screens changed after a release. See Demo specs and CI.

For web apps and app reviews it covers recording only the browser page with a drawn address bar (browser, url_contains), handing sign-in and consent to you (take_handoff), leaving slow server work out of the video (take_wait), exports under an upload limit (max_mb) and the platform check (review_check). See Record a platform review video.

And music: music_generate adds instrumental background music (checking the allowance with dry_run before a playlist), music_add uses your own track or lowers the music under narration, and composition_create builds a long-form study-with-me video with a timer, chimes and chapters. See Background music and Make a study-with-me video.

The guide inside the app

DemoTake also explains itself over MCP, so an agent without the skill can still find its way:

  • The server's connection instructions summarize the workflow and say which edition is running.
  • The demotake_guide tool returns the full workflow guide from the app itself, tailored to your edition, including browser takes, hand-offs and platform review presets. DemoTake Studio's guide, for example, tells the agent to operate the app with its own tools and report each step with take_log_event.
  • The same guide is available as the MCP resource demotake://guide, and the record_feature prompt starts a complete "record this feature" request.

If your agent seems unsure, ask it to "call demotake_guide first". It's read-only and cheap.

Tip:

The guide also covers privacy: mark secret input as secure so it's never logged and is masked in renders. The record_feature prompt adds that agents should never type real secrets you didn't provide. Snapshots an agent receives have detected secrets blurred.

Where to go next

Was this page helpful?