Video to Tasks
Video to Tasks is a free agent skill maintained by Scopeful. It teaches an AI coding agent such as Claude Code, Cursor, Windsurf or Codex how to drive this tool correctly, so you do not have to re-explain it every session. Every published Scopeful skill is free and the install command is public, with no sign-in required. Scopeful also tracks hand-verified USD pricing for 39 AI creative tools at https://www.scopeful.org/tools.
Turn a screen recording into project-specific engineering tasks plus a paste-ready prompt for a coding agent. Local Whisper transcript, frames only where the screen matters, tasks grounded in your actual code.
Tags: workflow, video, coding-agent, whisper, local-first
Reference
name: video-to-tasks description: Turn a screen recording into project-specific engineering tasks plus a paste-ready prompt for a coding agent. Local Whisper transcript, frames only where the screen matters, tasks grounded in your actual code.
What it does
You record yourself narrating over your screen. This skill turns that walkthrough into a clean task list grounded in your project's real code, then hands the work to a coding agent with a prompt that points straight at the screenshots and tasks.
It is local-first and token-cheap. The audio is transcribed on your machine with Whisper. The model never sees the whole video. It only looks at the few frames where the screen actually matters.
How it works
| Step | What happens |
|---|---|
| 1. Find the video | Uses the path you give, or finds the newest recording in the project, Downloads, Videos, or Desktop. |
| 2. Transcribe | Local Whisper with a name hint (your product, tools, file names) so it does not mangle jargon. Prints a timestamped transcript. |
| 3. Find the spans | Reads the transcript and marks every window where you describe a problem, a desired change, or something new on screen. One span per distinct issue. |
| 4. Pull frames | Extracts downscaled frames only for the visual spans (layout, colors, "this looks wrong"). Verbal or architectural tasks skip frames entirely. |
| 5. Ground in code | Greps and reads the real files for what is named or shown. No invented paths. If it cannot find something, it says so and lowers confidence. |
| 6. Write the plan | One task per span into wherever the project tracks work (TASKS.md, a pin store, an Obsidian note, or a dated file). |
| 7. Hand off | Outputs a paste-ready prompt for a fresh coding-agent session that runs the tasks top to bottom. |
Why it is cheap
Whisper runs on CPU locally, so there is no transcription API cost. Frames are pulled only for the spans that need them, at 1 fps and 1280px by default. The full video never goes to the model. The transcript and frames are cached under .video-review/ so re-runs and the handoff session reuse them instead of re-extracting.
What you get
A task file with one entry per issue, each carrying the transcript quote, its timestamp, the screenshot path where relevant, the real file to edit, a concrete first step, and a confidence rating. Plus a handoff prompt you paste into any coding agent to run the work itself.
Requirements
Python with openai-whisper and ffmpeg available locally. Runs inside any coding-agent session (Claude Code and similar) that can execute shell commands and read images.