ProducerDesktop app for WindowsVersion 0.1.0Built on HyperFrames and Claude CodeSteward: [name]
Drop a screen recording. Get a finished demo video.
Producer is a Windows desktop app that turns a raw screen recording into a narrated, branded demo video. It transcribes the recording, has Claude write a value-first script against what was said and shown, records studio narration on your machine, builds a motion-designed composition in one of eight visual templates, verifies every scene and seam, renders a 1080p MP4 locally, and packages it with captions, a thumbnail and the script. A Studio lets you direct changes by prompt or by hand.
[9]
Changes shipped this quarter
[78]%
Started as a member’s problem
[5]
Open requests in the queue
[4] days
Request to decision, median
Sample values. Live figures for this product, published monthly. How the scorecard works.
Who it is for
Product marketers, sales engineers, founders, customer success and internal communications teams who have a recording and a message. Agencies producing demo videos for clients in a consistent house look.
Who it is not for
Talking-head or live-action editing. Multi-camera or colour-graded productions. Anyone without a Windows PC and a Claude subscription, for now.
Why Producer exists
A screen recording is evidence, not a video.
Someone still has to decide what the story is, write narration that matches what is on screen, cut around the loading spinners, design the frames, time the callouts to the words, and render something people will watch. That is a day of work for an editor, and it is why most product walkthroughs are twelve minutes of unedited screen capture. Producer does the editor’s job with a person at three gates: the brief, the script, and the Studio. Everything between the gates is automatic and verified.
The footage is the source of truth
Every number, label and claim in the narration must be visible in the recording. Claude reads the transcript and the actual frames, grabs full-resolution stills to check a figure, and audits every footage window frame by frame for blank or loading screens before it commits to it. Nothing is invented.
Narration drives timing
You lock the script, the app records each line, and the real length of each spoken line sets the length of its scene. Callouts land on the spoken word, never at the top of the scene.
Verification is built in, not optional
A build is not done until the HyperFrames lint and check pass, the seam gate confirms every transition, and midpoint snapshots of the real composition have been captured for you to look at. If the check fails, Claude fixes it and runs it again.
Seven steps, three gates
Every project moves through seven steps in the left rail. Each step unlocks the next. The brief, the script and the Studio are yours. Everything between them is automatic and verified.
Drop an MP4, MOV, MKV, WebM or M4V, or click to browse. Producer copies it into a project folder, probes duration, resolution and frame rate, extracts a thumbnail, and builds a timestamped contact sheet with one frame every few seconds. That sheet is what Claude looks at first when planning the story.
Transcription runs locally. Producer uses whisper.cpp, which it downloads and installs for you, or Python Whisper with GPU acceleration if it finds one on the machine. The result is a timestamped transcript alongside the frames, so the script is written against the words as well as the pictures.
The brief is the contract Claude writes against. One clear message beats a list of features.
| Field | What it sets |
|---|---|
| Message | One sentence. It is the spine of the script and lands by scene two. |
| Audience | Internal staff, leadership, prospects. Free text. |
| Length target | 60 to 90 seconds, 2 to 3 minutes, 3 to 4 minutes, 4 to 6 minutes, or 6 to 8 minutes. |
| Tone | Free text. The default is confident, plain-spoken, internal walkthrough. |
| Template | One of eight visual looks, with a live preview of each. Colours and fonts start from the template’s defaults and can be edited freely. |
| Brand | Organisation name, canvas, accent and ink colours, headline and label fonts from a bundled render-safe set, and a logo. |
| Voice and speed | Eleven narration voices with a preview button that reads your message aloud in the chosen voice. |
| Notes for Claude | Facts to include, things to avoid, names not to mention. |
| Captions | On by default. |
Claude reads the brief, the transcript and the frames, then writes a script of eight to twelve scenes. Each scene has a title, a three-to-six word on-screen headline, twenty-five to seventy words of narration, the footage windows in source seconds that show what the narration describes, and a note of what else is on screen: callouts, count-ups, chips, diagrams, with exact values.
The script follows a value-first spine: a hook in the viewer’s language, the message by scene two, and every later scene as evidence. Narration is written the way it should be spoken, with product numerals spelled out and homographs rephrased so the voice reads them correctly. The whole run is visible in the Claude panel while it works.
You can edit any scene by hand, add scenes, or re-draft the whole script. Then record. Each line is voiced on your machine by the Kokoro model, and the real length of every recording is what the build uses for scene timing. Re-voice a single line after an edit without touching the rest.
Using the locked script and the recorded narration, Claude authors one HyperFrames scene per line by copying the chosen template’s approved reference scenes and changing only content, footage windows and beat times. It crops browser chrome out of the footage, holds a still where a screen was on screen for under three seconds, places callouts on the spoken beat, stamps velocity-matched seams between scenes from a written ledger, and runs lint and the full check until the check passes.
When it passes, the app captures between four and twelve midpoint frames from the real composition, so you see what the render will look like before rendering. If the check fails, one button sends the log back to Claude to fix.
Studio is HyperFrames Studio served locally inside the app: the timeline, element selection and manual controls. Below it is a prompt box. Describe a change, for example “make the headline in scene 3 red and start it a second later,” and Claude edits the composition with your current selection attached, re-checks it, and Studio reloads. Narration changes are script edits, and the app tells you which lines to re-voice.
Rebuilding from the script warns you first, because manual Studio edits would be overwritten.
Rendering runs locally in headless Chrome and FFmpeg with a configurable number of parallel workers. Two presets: a fast preview at lower bitrate for review, and the final delivery at 1080p, high-bitrate H.264 with AAC audio. A three-minute video takes about four minutes on a fast laptop. Every render is kept in the project with its duration and size, and plays back inside the app.
Bundle the MP4 with a captions file generated from the narration on the same timeline, a thumbnail, the script as Markdown, a README, and optionally the editable project source without the recording, so anyone with Producer can re-open and re-edit it. Choose a destination folder and the package is zipped for sharing.
The eight looks
Every template ships three approved reference scenes: a cold-open hook, a footage scene with chapter tag, KPI count-up cards, chips and a punch-in, and a close with a flow diagram, closing line and logo plate. Claude copies them and changes only content, footage windows and beat times, so the layout never drifts and every video in a workspace looks like it came from the same studio.
| Template | Aspect | Character |
|---|---|---|
| Studio Dark | 16:9 | Navy canvas, footage in a framed window, a story column of cards. The house look, with card, chip, panel and diagram variants. |
| Editorial Light | 16:9 | Warm paper canvas, ink type, strong rules. Calm and print-like. |
| Cinema Full-Bleed | 16:9 | Footage fills the frame under a dark scrim; headlines and lower-thirds ride on top. |
| Social Vertical | 9:16 | For LinkedIn, Reels and Shorts: footage on top, big type and stacked callouts below. |
| Boardroom Slate | 16:9 | Slate canvas, split screen: large numerals on the left, footage on the right. |
| Sunrise Warm | 16:9 | Deep plum canvas with warm glows; headline on top, footage centred, chips beneath. |
| Blueprint Grid | 16:9 | Light technical grid, mono annotations with leader lines. For engineering audiences. |
| Minimal Mono | 16:9 | Black and white, hairline frames, one enormous condensed headline. |
All templates share the same rules: callouts are content-sized cards, never a panel stretched to fill a column; type is video scale; browser chrome is cropped; beats land on the words.
Narration voices
Eleven voices, all generated on your machine.
| Accent | Voices |
|---|---|
| American, female | Heart, Bella, Nova, Sky, Sarah |
| British, female | Emma, Isabella |
| American, male | Adam, Michael |
| British, male | George, Lewis |
Speed is adjustable per project. The default is Heart at 0.95×. Voice files are recorded when you say so and never edited by the agent.
The Claude panel
Every run is visible as it happens.
The session and model, each file read and written, each command run, and the text Claude writes as it decides. The project keeps a history of every run with its summary and its cost, so you know what the script, the build and each edit cost you.
Unattended production. Producer can run the entire pipeline without the interface: create the project from a recording and a brief, transcribe, script, voice, build, check, and fix up to a set number of rounds, writing a stage-by-stage report with timings and costs. A built-in self-test exercises the whole chain on a sample recording. The app can also be launched straight into a given project at a given step.
Setup
Setup checks the machine and fixes what it can with one click.
- Claude Code found and signed in, using your own Anthropic subscription. Install and sign-in actions open a terminal and walk you through it.
- FFmpeg and the HyperFrames engine, both bundled.
- The render browser, a one-time download of about 160 MB.
- The production skills Claude needs, installed into your Claude skills folder.
- A transcription engine: whisper.cpp installed automatically, or Python Whisper detected.
- The narration voice model, about 330 MB on first use, cached locally.
- A projects folder.
Settings
Defaults for every production.
The default look, brand colours, fonts and logo, default voice and speed, the Claude Code path and an optional model override, the Python path, the transcription engine and Whisper model, the number of render workers, and the projects folder. Each project can override brand and voice in its brief.
What runs where
On your machine: transcription, narration, composition, rendering, packaging, and all project files. Projects live in a folder you choose, one folder per video, with the recording, transcript, script, narration, compositions, renders and packages inside.
Through your Claude subscription: script writing, composition building, fixing and prompt edits, via Claude Code. Producer never holds an API key. There is no billing, credit or metering inside the app.
Requirements
Producer’s requirements are heavier than the other three products, so here they are plainly.
| Platform | Windows x64, per-user install, no admin rights required |
| AI | A Claude Pro, Max or Team subscription with Claude Code installed. Producer never holds an API key. |
| Disk | About 1.6 GB for the app, plus about 490 MB for the Whisper model, 330 MB for the voice model and 160 MB for the render browser, downloaded on first use |
| Rendering | CPU-bound. A three-minute video takes about four minutes on a fast laptop. |
| Input | One screen recording: MP4, MOV, MKV, WebM or M4V |
| Output | 1080p H.264 MP4 with AAC audio, SRT captions, thumbnail, script, optional editable source |
What it deliberately does not do
- It never invents a number, label or claim. If it is not visible in the footage, it is not in the video.
- It never regenerates narration on its own. Voice files are recorded when you say so and never edited by the agent.
- It never renders from the agent. Rendering is the app’s job and happens when you press the button.
- It never quotes the transcript verbatim. The transcript is evidence for a written script, not a subtitle track.
- It never asks questions mid-run. Claude works non-interactively, makes reasonable decisions, states them in its summary, and finishes.
At a glance
| Steps | 7: Source, Brief, Script & voice, Build, Studio, Render, Package |
| Templates | 8 looks, 16:9 and 9:16 |
| Voices | 11, generated locally |
| Length targets | 60–90 s to 6–8 min |
| Script | 8–12 scenes, 25–70 words each, footage windows per scene |
| Verification | Lint, check, seam gate, midpoint snapshots, automatic fix rounds |
| Render | Local headless Chrome and FFmpeg, preview and final presets, parallel workers |
| AI | Claude Code on your own Anthropic subscription |
The charter
What it holds, and what it will never become.
Every accepted request for Producer cites one of these lines. Every decline cites the line it would break. The steward reviews the charter with the community once a quarter.
The opinions it holds
- The footage is the source of truth. Every number, label and claim in the narration must be visible in the recording. Nothing is invented.
- Narration drives timing. The real length of each recorded line sets the length of its scene. Callouts land on the spoken word, never at the top of the scene.
- A build is not done until the check passes. Lint, check, the seam gate and midpoint snapshots, every time. If it fails, the agent fixes it and runs it again.
- Three gates belong to the person. The brief, the locked script, and the Studio. Everything between them is automatic.
- The agent never renders and never touches narration. Rendering and voice recording are the app’s, and happen when a person presses the button.
- Templates are copied, not reinvented. Every scene starts from an approved reference scene so every video in a workspace looks like one house.
What it will never become
- A general video editor. No timeline for live-action, multi-cam or grading. The input is a screen recording and the output is a demo.
- An avatar or stock-footage generator. Every frame comes from your recording or the template. No synthetic presenters, no generated b-roll.
- A video host or analytics platform. The deliverable is a file. Publish distributes it and measures it.
- A cloud service that holds your recording. Transcription, narration and rendering run on your machine. Claude sees the transcript and selected frames, nothing more.
What members can ask for
In scope
- A template look
- A narration voice
- A package output such as a vertical cut
- An input format
- A verification check
- A brief field
- A Studio control
Out of scope
- A live-action editing timeline
- AI presenters
- Generated footage
- An agent that renders or re-voices on its own
How it fits the family
The video maker. Publish distributes what it produces. Velocity uses the videos in sequences and on landing pages. Production tasks are tracked in LevelUp. Workflow handoffs, not built integrations, until the changelog says otherwise.
The Producer queue · live
What members asked for in Producer, and what we did.
| The problem | Raised by | Status | Decided in | Shipped |
|---|---|---|---|---|
| [Problem statement as the member wrote it] | [Member name] | Shipped | [3] days | [date] |
| [Problem statement as the member wrote it] | [Member name] | Building | [2] days | behind a flag |
| [Problem statement as the member wrote it] | [Member name] | Planned | [4] days | [sprint] |
| [Problem statement as the member wrote it]Declined by the steward: [the reason in one sentence, citing the charter line it breaks]. | [Member name] | Declined | [6] days | – |
Every request gets a decision within ten business days. Declines are public, with the reason. Shipped entries name the member who asked. The first entries appear as the founding cohort posts its first problems.
The routine
Per video, not per week. About ten minutes of your attention.
Gate 1 · The brief
Gate 2 · The script
Gate 3 · The Studio
Two to three minutes of video, about ten minutes of your attention.
Help, built in
Producer ships its own manual.
- Setup checks the machine and installs what is missing with one click.
- The Claude panel shows every run, and one button sends a failed check back to Claude to fix.
- A built-in self-test exercises the whole chain on a sample recording.
- Every changelog entry that changes Producer links to the help article it changed.
Producer is included in every membership.
Windows today. Needs a Windows PC and a Claude Pro, Max or Team subscription. Members get the download link and every new build from the changelog.
