Article 19 in the AgentKit v3 onboarding series. About a 9 minute read. This one is a creative workflow, not an install command. It connects to article 13 (subagents) and article 12 (picking a workflow by model).
I had Opus 5.5 make anime from a bare one-line prompt, then pushed it through five stages to produce the series FRONTIER (episode 01 is up on YouTube, episode 02 was still rendering when I wrote the original post). This article pulls out the reusable parts. It is not a promise that "anyone can make great anime with one command".
1/ The toolkit
| Tool | Role |
|---|---|
| Codex (the 20 USD plan is the minimum) | Character design, backgrounds, storyboards |
ffmpeg |
Cutting, joining, encoding |
| HyperFrames | Editing with HTML + GSAP, with a web UI for manual tweaks |
| Gemini 3.8 TTS | Voice-over |
| Gemini Omni Flash | Short shots of about 3 seconds: close-ups, lip movement, expressions |
| MiniMax H3 | Close-ups and audio-driven lipsync (open weights, can run locally) |
| ElevenLabs | SFX and supporting audio |
The model names are what I wrote in the original post at that time. Names and pricing change fast, so double-check.
2/ The five stages
01. Test the raw ability. Always start with a bare prompt to see how far the model gets on its own before you fill in the gaps. My prompt, roughly: use codex to create the art, then build a 10 to 15 second anime shot with HyperFrames. The result was fairly decent, but Opus let basic mistakes slip through, like deformed fingers, without noticing them itself.
02. Understand the limits and give hints. I showed Opus ufotable fight scenes to learn from. It still could not get there, and criticizing the model does not make it better. What helps is giving it new tools: Three.js and shaders for depth and effects, and splitting the background into layers with a depth model running locally, so when the camera moves the layers have depth.
QA subagents. At first I had to babysit constantly. Once I understood, I set a quality standard, listed Opus's common mistakes, and created a team of "QA" subagents. Every time the main session produces something new (an image, a video, SFX), it spawns the QA team to inspect it immediately, instead of waiting until the end to find the mistake and re-render everything. Rendering takes a long time, so this really saves effort.
03. Optimize, dogfood, package. 2D mesh animation (Spine 2D or DragonBones style) improved clearly but costs tokens. The other direction: use AI video only for short shots of about 3 seconds or to edit existing video, and don't make it carry long scenes. After that, have Opus re-read the whole session to extract learnings and mistakes, then package them into a skill so next time is cheaper and faster.
04. Longer videos. What humans keep is the script and the aesthetic taste. I write an outline, ask ChatGPT to develop it into a 3 to 5 episode series, review the script myself, then Opus reads the script, codex creates the characters, and storyboards are made. The hardest part is keeping characters consistent: Opus creates two files by itself, continuity.json (character state per scene: appearance, accessories, injuries, outfit) and story-so-far.md (a summary of the story so far). The QA subagents rely on these two files to catch detail errors.
05. Run locally, near 0 cost. MiniMax H3 is open weights, its API is about 40% cheaper than Gemini Omni by my estimate, and running it locally costs only electricity. On my machine, a 4 second 720p clip takes about 5.3 minutes. It accepts both audio and a reference image, so lipsync matches the TTS dialogue line. Free Vietnamese TTS is still being tried, not solved yet.
Another snag: Opus 5.5 has no ears, so SFX is hard. My way around it was to ask it directly what it needs, and it suggested sending it references to learn from, plus using free SFX libraries.
3/ How to reuse this (you don't need my exact setup)
Since the original post does not give fixed commands, I only list the steps as principles:
Test a bare prompt first and write down the common mistakes.
Instead of criticizing, add tools: depth, shaders, Three.js, mesh.
Write the QA standard and spawn QA subagents after every new asset.
Save character state to files.
Use AI video for short shots and close-ups.
Dogfood the session, then package it into a skill.
The AgentKit kit has skills related to this, for example ak-motion-video (motion graphics video with HyperFrames, with voice-over, captions, SFX, music) and ak-filmmaking (long films, consistent characters, storyboards, cost). The skill that packages the anime workflow I mentioned in the original post is a separate one (anime-creator is in my own skill catalog); I did not check its commands in this article so I do not write its syntax here. Read the skill's SKILL.md before running it.
4/ Status
| Part | Label |
|---|---|
| The bare prompt flow, tool hints, QA subagents, continuity files | ✅ I did this for real, and it produced episode 01 |
| Gemini Omni Flash for short shots | ✅ I used it, API is billed |
| MiniMax H3 running locally | 🧪 tried it: 4 seconds 720p takes about 5.3 minutes on my machine, depends on your hardware |
| LTX-2.5, ZeroTTS, OmniVoice for a 0-cost pipeline | 🧪 testing, no conclusion yet |
| Free Vietnamese TTS with enough emotion | 🗺️ not solved yet |
The relative costs in the original post (Omni = 100, MiniMax H3 API = 60, local = 0) are my own normalized figures, not an official price list.
Wrap-up
1/ Start with a bare prompt to learn the model's real limits.
2/ To make the model better, give it tools, don't criticize.
3/ QA subagents check each asset as soon as it is created, and state files keep characters consistent across episodes.
4/ AI video is for short shots only. The script and the aesthetic taste are still the human's job.
Sources
How I make anime with Opus 5.5 (24/10/2026, updated 25/10/2026)