Topic2Reel
Type a topic. Get a ready-to-post Instagram Reel: animated 9:16 scenes, Hindi voiceover, music and a caption with hashtags.
It runs on your own machine, needs no API key, and ships as a Claude Code plugin plus a plain command-line tool.
topic ──► spec.json (written by Claude, or any LLM, from prompts/topic_to_spec.md)
──► engine/topic2reel.py ──► reel.mp4 + reel.caption.txt
What you get
| |
|---|
| Output | 1080×1920, 30 fps MP4, about 30–45 s, plus a caption file |
| Visuals | 7 scene types (hook, cards, myth/fact, stat, steps, compare, call to action) drawn in code, no stock footage |
| Voice | Hindi voiceover with Piper (optional; without it you get music only) |
| Cost | Free. About 2 minutes of CPU time for a 35 s reel |
Use it as a Claude Code plugin
/plugin marketplace add sahibsingh/Instagram-topic2Reel
/plugin install topic2reel@topic2reel
/topic2reel:reel study tips for board exams
These commands follow Claude Code's plugin documentation as I know it. This repository's plugin has not been loaded in a
live Claude Code session yet; the engine, the tests and the page are what has been run. To try it from a local folder:
git clone https://github.com/sahibsingh/Instagram-topic2Reel.git && claude --plugin-dir ./Instagram-topic2Reel.
Claude then runs the workflow in skills/reel/SKILL.md: check the machine, write the spec, validate, preview, render, report.
Prompt builder and templates
The demo page (docs/index.html) has a prompt builder: pick a topic, length, tone, scene types, hook, call to action and voice, and the
prompt updates live. Two tabs: the slash-command form for Claude Code with the plugin, and a longer form for Claude in the cloud (it also
works without the plugin and has an Open in Claude button). Twelve templates (fitness, study, money, safety, career, ...) fill the
form for you. The builder writes a short KEY: value brief that skills/reel/SKILL.md understands (TOPIC, AUDIENCE, TONE, LENGTH,
SCENES, HOOK, CTA, VOICE, PRONUNCIATION, PREVIEW, SETUP, FACTS, AVOID). A test keeps the two in sync.
Use it from the command line
python3 engine/topic2reel.py --doctor # what is missing on this machine?
python3 engine/topic2reel.py examples/study_tips.json --validate
python3 engine/topic2reel.py examples/study_tips.json --preview out/prev # a few still frames, about 10 s
python3 engine/topic2reel.py examples/study_tips.json -o out/study.mp4 --voice piper --piper-model "$TOPIC2REEL_PIPER_MODEL"
python3 engine/topic2reel.py examples/study_tips.json -o out/silent.mp4 --voice none
python3 engine/topic2reel.py spec.json -o out/x.mp4 --voice dir --voice-dir ./my_voice # your own scene_1.wav ... scene_N.wav
From Python: from topic2reel import generate; generate("spec.json", "out/reel.mp4", voice="none").
Behind your own feature flag, ask your LLM for a spec with prompts/topic_to_spec.md as the system prompt, check it with
validate_spec, then call generate.
Install
pip install -r requirements.txt # numpy, piper-tts (GPL-3.0, see below)
npm ci && npx playwright install chromium # the renderer (package-lock.json is committed)
sudo apt install ffmpeg fonts-noto-color-emoji fonts-inter # or your OS equivalent
scripts/get_voice.sh hi_IN-pratham-medium # downloads a Hindi voice, prints the export line
export TOPIC2REEL_PIPER_MODEL=~/.topic2reel/voices/hi_IN-pratham-medium.onnx
python3 engine/topic2reel.py --doctor
scripts/get_voice.sh uses Piper's own downloader and has not been run on the machine that built this repo (no access to
Hugging Face there). Voice models are not stored in this repository (about 63 MB each).
The spec
{ "caption", "hashtags", "scenes": [4 to 8 scenes] }, schema in schema/reel-spec.schema.json, examples in examples/.
Every scene has type, headline (1–2 short lines), voice, and optional highlight and caption.
type | extra fields | looks like |
|---|
hook | emoji, stamp, badge | big emoji sticker with a red cross or green tick |
cards | items[{emoji,label}] | grid of icon cards popping in |
myth_fact | emoji, myth, fact | myth bubble flips into a green fact bubble |
stat | value, emoji, label, sub | big number, counts up if numeric |
steps | items[{text,emoji}] | checklist that ticks off |
compare | left, right | the wrong way vs the better way |
cta | emoji, primary, secondary | save / share buttons and hearts |
A scene lasts as long as its voice line plus about one second; without voice it lasts 4.4 s.
Limits
- On-screen text must be Roman script (Hinglish is fine). No Devanagari font is bundled, so the validator rejects it;
Devanagari is only for
voice.
- Visuals are emoji and shapes, so they fit any topic but look simpler than hand-drawn scenes.
- Voice quality depends on the Piper voice, and English loanwords can be mispronounced. Re-spell them in Devanagari in
voice.
- The facts come from the LLM that writes the spec. Read health, money and legal topics before posting.
- The music is a small synthesized beat. Swap in a trending audio when you upload.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
