Prompt & Done · REELKIT

The kit to build your own AI reels

A free, open-source (MIT) starter kit: Claude Code researches and writes the script, a local AI voice reads it, Whisper checks every word, and a 3D scene renders frame by frame into a finished vertical video.

v1, 61 KB. Tested 4 Nov 2026 on an Apple silicon Mac: from a fresh unzip, the 11-second sample video rendered and passed all 16 automated checks in 39 seconds (models already downloaded).

The pipeline, in 7 stages

Claude Code follows one skill file (.claude/skills/make-reel/SKILL.md) through every stage. Each stage has a pass/fail gate before the next one starts.

1 Research

Searches what your audience is watching this week, proves demand with real numbers, ranks 5-8 ideas and picks the strongest.

Tool: Claude Code with web search

2 Test it and fact-check it

Runs the tool for real and saves the input and output, then builds a claim table where every fact has a primary source.

Tool: Claude Code

3 Script

One spoken paragraph of 75-95 words: the result in the first 1-2 seconds, then the proof, the payoff and one call to action.

Tool: Claude Code + the make-reel skill

4 Voice and word check

A local AI voice reads every line several times. Whisper transcribes each take and keeps only exact word matches, then times every word for the captions.

Tool: Kokoro-82M, faster-whisper

5 3D scene

A three.js world (an original voxel robot on a floating island to start) acts out the script with one smooth camera and word-synced captions. Every frame is a pure function of time, so renders repeat exactly.

Tool: three.js

6 Render and mix

Headless Chromium draws all frames at 30 fps; an original music bed and sound effects are ducked under the voice, mixed to -14 LUFS and encoded as a 1080x1920 MP4.

Tool: Playwright, FFmpeg

7 QA, then your OK

16 automated checks on the final file (size, codec, loudness, true peak, black or flash frames, hard cuts, and Whisper hearing the script exactly) plus a contact sheet. Then you watch it and decide. The kit never posts.

Tool: qa.py (FFmpeg + Whisper), you

A The first prompt

After setup, open a terminal in the reelkit folder, run claude, and paste this. Replace <TOPIC>. More prompts (research only, revise, a batch run while you sleep) are in PROMPTS.md.

Make one reel about <TOPIC>. Follow CLAUDE.md and the make-reel skill end to end: research and prove demand, test it for real, build the claim table, write the script, voice, ASR gate, build the scene, render, mix, run QA and fix what fails, then package it. Don't stop to ask me unless you are blocked. Look at the contact sheets yourself before you call it done. Do not post anything. When you finish, give me the MP4 path, its duration, the QA result, and anything you could not verify.

B Your channel file: CLAUDE.md

Claude Code reads this at the start of every session. Fill in the <...> parts once: your channel, voice, hook rules, CTA, accuracy rules, quality bar, and "never post without my OK". It is already in the kit.

# Channel instructions (fill this in)

Claude Code reads this file at the start of every session in this folder. Replace every `<...>`. Keep it short: rules here apply to every video. The step-by-step production process lives in `.claude/skills/make-reel/SKILL.md`.

## The channel

| | |
|---|---|
| Name | `<Channel name>` |
| Promise (one line) | `<e.g. "One AI tool a day that saves you an hour. Shown, not told.">` |
| Audience | `<who, what they already know, what they want: e.g. "people who use ChatGPT or Claude at work and want to do more with it">` |
| Niche to research | `<topics in scope, e.g. "AI tools, agents, coding assistants, open-source AI you can run">` |
| Out of scope | `<e.g. "crypto, finance tips, politics, celebrity news">` |
| Handles | `<TikTok @...> / <Instagram @...> / <YouTube @...>` |
| Length | 25-35 s by default (75-95 spoken words) |

## Voice and tone

- `<e.g. "Calm, confident, a little dry. Talks like a friend showing you something on their laptop.">`
- Write for speech: short clauses, contractions, one idea per sentence, concrete examples.
- Never: "Hey guys", "Welcome back", "In today's video", "game-changing", "revolutionary", "insane".
- TTS voice: `<Kokoro voice id, e.g. af_heart or am_michael>` (set `voice.kokoro_voice` in each script.json).

## Hook rules

- The first 1-2 seconds show the result or the surprising capability. Never open on a logo, a greeting or the tool's name.
- Speech starts at 0.0 s. Frame 0 is the most striking image in the video.
- Proof on screen by about 3 s.

## Call to action

- One CTA, after the payoff, while the scene is still moving.
- Default: `<"Comment KEYWORD and I'll send you the link." or "Follow for one AI tool a day.">`
- A keyword is short, distinctive and not a common word. Only promise what exists and works.

## Accuracy rules

- Fresh research before every script. Never script from memory.
- Every claim (names, prices, plans, licences, dates, capabilities, numbers) has a primary source or our own test in the reel's claim table.
- Never repeat a creator's claim without checking it. Never invent numbers or engagement.
- Scope claims precisely instead of adding spoken disclaimers. Put prerequisites (paid plan, hardware) in the post caption.
- Label illustrative scenes. Show the real input and the real output whenever possible.
- Never clone or imitate a real person's voice without their written consent. Never use a trademarked mascot or logo as a character.

## Quality bar

- 1080x1920, 30 fps, H.264, AAC, -14 LUFS, true peak <= -1.5 dBTP.
- `make qa` passes, and you looked at the contact sheet and the key frames yourself.
- Captions: max 4 words, word-synced, nothing overlapping the hero.
- One continuous 3D world per video, at most two hard cuts, something new every 2-4 s.

## Log of past videos (avoid repeats)

| Date | Slug | Topic | Notes |
|---|---|---|---|
| `<yyyy-mm-dd>` | `<slug>` | `<topic>` | `<what worked, what didn't>` |

## Permissions

- **Never post, schedule, upload or send anything without my explicit OK for that specific video.** Deliver the finished file and the package; I watch it and decide.
- Do not spend money (paid APIs, subscriptions) without asking. `<adjust if you allow a budget, e.g. "up to $2 of API calls per video">`
- Work inside this folder. Don't delete anything outside `reels/<slug>/`.

C Setup on a Mac (about 10 minutes)

  1. Get Claude Code on a paid Claude plan and log in once.
  2. Install FFmpeg, Node 18+ and uv with Homebrew.
  3. Unzip the kit and run make setup, then make all to render the sample.
  4. Open example/final.mp4 and example/qa.json, fill in CLAUDE.md, and paste the first prompt.
brew install ffmpeg node uv
curl -fsSL https://claude.ai/install.sh | bash   # Claude Code; run `claude` once to log in
unzip reelkit-v1.zip && cd reelkit
make setup     # Python env, Node packages, headless Chromium (~1.7 GB)
make doctor    # every line should say ok
make all       # renders the sample video: example/final.mp4

Needs about 2.5 GB of disk; the first run downloads the voice and Whisper models (about 0.8 GB). Linux and Windows (WSL) steps are in the README; we have only tested on macOS so far.

D The stack and licenses

PartJobLicense
Claude Coderuns the whole processneeds a paid Claude plan
Kokoro-82Mtext to speech, runs on CPUApache-2.0
faster-whisper + Whisper small.enspeech to text for the word check and QAMIT
three.jsthe 3D sceneMIT
Playwright + Chromiumframe-by-frame renderingApache-2.0 / BSD-style
FFmpegmixing, loudness, encodingLGPL/GPL
espeak-ng (bundled)pronunciation fallback for KokoroGPL-3.0
Chatterbox (optional)clone your own voiceMIT
Replybeam (optional)sends the link to everyone who comments your keyword on Instagramfree plan; Pro $9/month
The kit itselfskill, scene, scripts, docsMIT

The music and sound effects are synthesized from plain math by the kit, so there is nothing to license. Every tool except Claude Code is free and runs on your computer.

F Send the link to everyone who comments

Most reels end with "Comment KEYWORD and I'll send you the link." Sending it by hand doesn't scale. Replybeam does it for you on Instagram: when someone comments your keyword, it posts a short public reply and DMs them your link as a button, and tracks the clicks.

Disclosure: Replybeam is made by the team behind Prompt & Done.

E Honest notes

Voice in the reel: Built with Fish Audio.