# Prompt: a review video of a tool, made with the tool (HyperFrames) This is the whole process behind my HyperFrames review video, as it ended up after seven versions, written as one prompt for your own AI coding agent. What's taken out: names, usernames, file paths, account, voice and asset IDs, and credit balances. Replace everything in [square brackets] with your own. ## Before you paste it 1. Turn HyperFrames telemetry off before you run anything else. Set the environment variable HYPERFRAMES_NO_TELEMETRY=1 in your user environment, so every new terminal and every command your agent runs sees it (a one-off "set" in a single terminal doesn't last). Then check it with `npx hyperframes telemetry status`. 2. Install Node 22 or newer and FFmpeg. 3. An AI coding agent that can run commands on your computer. I used Claude in the Claude Code desktop app. Install the HyperFrames skills from a terminal: `claude plugin marketplace add heygen-com/hyperframes`, then `claude plugin install hyperframes@hyperframes`. The install may warn that some files arrived as Git LFS pointer files; mine did, and the skills still loaded. 4. `npx hyperframes doctor` lists what's missing. Only some of it is needed to render. 5. Optional: a voice, avatar, music and sound-effect service if you want narration, an on-camera host and a mix. Set a spending limit you're comfortable with first. Know what you're starting. The prompt records your screen. Your agent looks at frames of those recordings, which sends the images to your AI provider. And if you use paid voice, avatar or music services, it spends money, within the limit you set. ## The prompt In Claude Code the first line is `/hyperframes:hyperframes`; other agents may call the skill `/hyperframes`. ```text /hyperframes:hyperframes You're making a review video of [TOOL], and the video is made with HyperFrames: the screens in it are recordings of you installing, testing and using [TOOL] (and HyperFrames itself, if HyperFrames is the tool). Work in [PROJECT FOLDER]. Run the HyperFrames CLI the way its skill says to. Ask me only when you're blocked: a creative call I need to make, a missing input, any spending or upload, installing anything outside the project folder, or signing in to or creating any account. Decide everything else yourself and log it. 1. Ground rules - Publish nothing. You deliver files; I review and upload them. - Never type, show or record a password, key, token or payment detail. - Spending: before every paid generation, run an estimate, tell me the cost and the running total against [CREDIT CEILING], and generate one take at a time. If you can't estimate a cost, or no ceiling is set, ask me first. Log every paid run (date, model, units, cost, what it's for). Check what comes back against what you asked for (length, takes, volume) before using it. - Every number and every claim about a tool in the script, on screen or in captions goes in a FACTS.md table with its source and the date you checked it. No row, no claim. - Confirm telemetry is off (npx hyperframes telemetry status) before the first HyperFrames command. 2. Record the work - Before each session I turn on Do Not Disturb and close anything private. Record one display only, with FFmpeg at 24 fps and a keyframe every second (-g 24), one recording per session. Start each session with a 5-second test clip and check it for anything private before the real take. - Command output often prints file paths with my operating-system username in them. Treat every frame as private until checked: use only stills or clips you have redacted, pixelate the username wherever it appears, crop out the taskbar and any sidebar with account details, and keep account pages, emails and notifications off screen. - Keep a CAPTURE-LOG.md with a timestamp, the command and what happened, especially what broke. The "what broke" chapter is cut from this log. 3. Test the tool for real - Install it, run its health check, make a starter project, lint it, check it, snapshot it, preview it and render it. Time the renders. Note every error, warning and surprise exactly as printed. - Read its docs and FAQ for what it costs, its license, what it sends home (telemetry) and what needs an account. Date each fact. 4. Script - About [LENGTH] minutes in five parts titled as questions: a cold open, what it is, how it works with an agent, what broke, and who it's for. - Write for the ear: one idea per sentence, 20 words or fewer, numbers said the way a person says them, exact figures on screen. - [IF YOU USE AN AI HOST: Say in the first 15 seconds that the host is an AI persona and that the video is made with AI.] Credit tools by name as facts. No affiliate links or sponsor claims unless they're real and disclosed on screen and in the description. - Show me the script and freeze it before any narration is generated. 5. Media (skip what I don't use) - Narration in [VOICE], one take per line, leveled to about -20 LUFS, with word timings from a local speech recognizer for the captions. - [Optional: an on-camera host from one approved still, lip-synced to the narration.] - [Optional: B-roll of the host at work, played slowed down where a line is about them.] - A music bed ducked under the voice. 6. Direct it; don't template it Follow the HyperFrames skill's film process: references/story.md, references/spec.md, references/motion.md, references/sound.md, references/captions.md, references/review.md and references/types/explainer.md. - Write SPEC.md before any scene: every look choice as a number, a BANNED line (start from the skill's "BANNED by default" list), and the frames you'll check. - A look built from one medium. Mine: an editor's pin board at night. The real screenshots are pinned prints with a small tilt and a contact shadow on a dark board with a faint dot grid. One amber marker draws a box around the line that matters and underlines a note's key words. A note in a display serif sits beside each print. - One camera per chapter travels from print to print and lands dead still. Hard punch-ins on long holds, after the note has had time to be read. No cross-fades between screens. - Lines with no screen are set as type (one line, a stack, a number that lands, a list, a ledger), and each part appears when the voice says it. - Captions on a scrim sized to the text, the spoken word in the accent color. - Sound: few and exact. Every sound has a cause on screen; repeated sounds rotate two or three takes; no synthesized tones or dings; level every sound file before mixing, and place each so its peak lands one frame after its picture. My kit: swishes on camera moves, a marker stroke per print, taps as type appears, a paper card on chapter titles, a swell into B-roll, one riser into the end card. 7. Check before you render, and after - npx hyperframes check --json until it has no errors; read every warning and fix it or say why it stays. - npx hyperframes snapshot --review, plus snapshots at the QA frames in SPEC.md. Describe each frame as a stranger would and fix anything that reads wrong. - After rendering: scan every frame for empty ones (measure your background's brightness first, so the threshold sits above it), run npx hyperframes audio --json for loudness and clipping, and master to about -14 LUFS with a true peak at or below -1 dBTP. - Write REVIEW.md: what you decided, what you measured, what nobody checked (you can't hear the mix), and the weakest shot. 8. Deliver - The master, an SRT of the captions, YouTube chapters, a description draft with the disclosures and a "how it was made" paragraph, and phone-size copies for review. - Then [N] cut-downs of 25 seconds or less in 1080x1920, each opening on a hook in its first second. ``` ## What to expect - Plan on several versions. Mine went from a rough cut to a templated look to the directed one above; the biggest jump came from writing SPEC.md and banning the defaults. - Check the screenshots yourself before anything goes public. Usernames show up in places you don't expect. - Generated sound effects can come back at very different volumes, some silent. Measure them before you mix. - Paid generations don't always return what you asked for. My music bed was billed for 285 seconds and came back 60 seconds long. Check the length before you build on it.