Open-source agent skill · MIT

Your coding agent
makes the episode.

okaypic-video is a skill for Claude Code, Codex and other agents: describe a story, and the agent writes the shot list, draws the cast, renders every take with MiniMax H3 — voices and sound included — and cuts the film with English or Chinese captions.

$npx skills add okaypic/okaypic-video-skill
View on GitHub
“GOBLIN CITY · Episode 1 · The Liquidation” · 3:52 · 28 shots rendered · about $6 in API calls

Workflow

From a prompt to a finished cut

The agent does the creative work and quality control; the scripts do the mechanical work. You approve the plan and the faces before anything is rendered.

01

Plan

Concept, tone, length and language with you. For a series, an outline you sign off first.

outline.md

02

Shot list

One shot per H3 clip: place, action beat by beat, spoken lines in quotes, the voice, the sound.

shots.json

03

Reference sheets

A sheet per character (close-ups + front / side / back) and per place — the faces every clip inherits.

cast.json → refs/sheets/

04

Takes

Two MiniMax H3 takes per shot, up to 18 in flight, a 5-frame contact sheet per take to pick from.

gen_clips.py

05

Cut

A local timeline editor: pick takes, trim, nudge captions while watching, write the Chinese captions.

editor.py

06

Export

1280×720, burned-in English or Chinese captions, per-clip levelling, −14 LUFS, your end card.

edit.py

Second skill · product ads

okaypic-ads: a product ad from photos and a sentence

The same workflow as Flow → Product ad, run by your agent: it looks at your product photos, asks what it still needs and what the video is for, proposes features, taglines and stories to pick from, casts one of six AI actors with their own voice, shows you every keyframe for approval, renders MiniMax H3 clips and cuts the ad in 9:16, 16:9 and 1:1.

$npx skills add okaypic/okaypic-ads-skill

A 20-second ad with three keyframes and two takes per clip is about $0.50.

“Lucky Bamboo” · 20 s lifestyle video · one photo and one sentence · $0.52

Editor

A real timeline, in your browser, from one Python file

python scripts/editor.py ep01 — no npm, no build step. Pick takes, trim on the timeline, drag captions, export EN / ZH.

The okaypic-video local editor: preview with captions, take inspector, clip track and English / Chinese caption tracks

Agents

Works where you already work

Claude Code

Install the skill and describe the video; Claude plans, renders and cuts it with the scripts.

Codex

Reads the same SKILL.md. Codex draws the reference sheets with its own built-in image model, so only the video clips use the okaypic API.

Any SKILL.md agent

Plain Python (stdlib only) and ffmpeg — the scripts also run on their own, without an agent.

Costs

Pay per second, see it first

10-second H3 clip, 768p$0.10
10-second H3 clip, 480p$0.07
Character or place sheet (GPT Image 2.5)$0.03
Character sheets drawn by Codex itselffree on okaypic
A 25-shot episode, two takes per shotabout $5–6
  • ✓ States the number of calls and the cost before every paid step, and waits for your OK.
  • ✓ gen_clips.py --quote prices only what is left; a resumed batch never pays again for finished or in-flight takes.
  • ✓ Every request carries a client_request_id, so a retried submit never charges twice. Failed renders are refunded.
  • ✓ Checks the balance first; if it runs out mid-batch it stops with a top-up link and re-running continues where it stopped.

Questions

What do I need?

+

An agent that reads SKILL.md (Claude Code, Codex and others), Python 3.9+, ffmpeg, and an okaypic API key. New accounts get $0.30 of free credit, enough for a few test clips.

How do I install it?

+

npx skills add okaypic/okaypic-video-skill, or clone github.com/okaypic/okaypic-video-skill into .claude/skills/okaypic-video (Claude Code) or ~/.codex/skills/okaypic-video (Codex).

What does an episode cost?

+

H3 video is $0.01 per second at 768p ($0.10 per 10-second clip) and $0.007 at 480p; reference sheets are $0.03. A 25-shot episode with two takes per shot is about $5–6. The skill quotes the cost before every paid step.

What if a batch stops halfway?

+

Re-run it. Finished takes are skipped, takes already submitted are polled instead of resubmitted, and only what is left is quoted and charged. A lost submit is resent with the same client_request_id, so it is never charged twice.

Can I use my own image model?

+

Yes. Agents with their own image tool (Codex has GPT Image built in) draw the reference sheets themselves; only the video clips go through the okaypic API.

Is it open source?

+

Yes, MIT-licensed. The scripts use only the Python standard library and ffmpeg and also run without an agent.

Make your first episode

Install the skill, get a key with $0.30 of free credit, and describe the story.