<- all tokdocs
2026.09.16@imjonathanacuna

Turning Claude Code Into a Video Generator: API Key, Pasted Seedance 2.0 Docs, and a One-Line Brief

AI
Watch on TikTok

View on TikTok ->

The "one prompt" is three pasted blocks: a Kie.ai API key, the full Seedance 2.0 API documentation, and a single sentence describing the video you want. Jonathan Acuña (channel name "Jonathan Acuña - Doctor AI") records his MacBook screen while setting up a Claude Code session and typing the prompt. The video ends before Claude produces anything, so the output is never shown.

The setup sequence, and where it did not go as described

The terminal shows him launch claude --permission-mode auto. Claude Code v2.1.259 starts on Haiku 4.5 under a Claude Max plan, and the status bar reads "manual mode on" with "auto mode unavailable" at the right. The Claude Code permission-modes doc explains this: when auto is requested but the model does not support it, the session silently starts in Manual. Haiku is listed as unsupported for auto mode on every provider.

He then types /effort ultracode. Claude Code rejects it: "Ultracode runs at xhigh effort, which [Haiku] doesn't support. Switch to an xhigh-capable model (Opus 4.7+, Sonnet 5)." He opens /model and picks Opus. The confirmation reads "Set model to Opus 5 (1M context)" and the status bar shows "effort high". He never re-runs the effort command, so the session runs at high effort, not ultracode. Per Claude Code issue #69653, ultracode is xhigh effort plus workflow orchestration, not a tier above max.

The prompt

The composed prompt in frames 22 through 24 reads, with pasted blocks collapsed by the terminal:

Kie API Key: Paste Here

Seedance 2.0 Documentation
[Pasted text #1 +720 lines]

I want you to create a 9:16 video with my heads[hot]
eating a hamburger. [Pasted text #2 +5 lines]

The last line is cut off at the screen edge; "headshot" is the most likely word. The second pasted block is five lines and its contents are not visible. He then clears the line and starts retyping "Video P..." as the clip ends.

The pattern: give Claude Code credentials, give it the API reference so it does not guess endpoints, then state the deliverable in plain language. Claude writes and runs the code that calls the API. The Kie.ai Seedance 2.0 doc confirms this is a two-call job: POST /api/v1/jobs/createTask with a Bearer token, then poll for the result. The API accepts text plus up to nine reference images, three reference videos, and three audio files, and lists 9:16 as a supported aspect ratio. Clips run 4 to 15 seconds.

The slash command list hints at a reusable skill

When he types / the autocomplete shows a custom /graphic-artist command whose description begins "Use when gen[erating]... kie.ai (Nano ...". The rest is off screen. He already has a saved skill that wraps Kie.ai generation. The video does not use it; it pastes the docs raw instead.

What Seedance 2.0 is

Seedance 2.0 is ByteDance's video model, released by the Seed team in February 2026. The official launch post describes four input modalities: text, image, audio, and video. Kie.ai resells API access to it. The video's title says "graphic artist", but the demo is video generation, not static graphics.

Claims I could not verify

The title says this makes Claude a "six figure graphic artist" and the narration says clients fire their graphic designer, their Higgsfield operator, and their content person within an hour of receiving the prompt. No client, invoice, or output is shown. Treat these as marketing. Higgsfield is real (an AI creative suite that offers Seedance 2.5 among other models), but nothing in the video shows it.

Key Takeaways

  • The prompt is API key + pasted API docs + one-sentence brief. Nothing else.
  • Pasting the full API reference is the part that matters; it keeps Claude from inventing endpoints.
  • Auto mode and ultracode both failed on Haiku 4.5. He switched to Opus 5, but the session ended at high effort, not ultracode.
  • The Kie.ai Seedance 2.0 API is a create-task-then-poll flow with Bearer auth and supports 9:16.
  • No generated output appears in the video. The "six figure" and "fire the team in an hour" claims are unverified.
  • A saved /graphic-artist skill exists on his machine, which is the repeatable version of this trick.

Resources

Published September 16, 2026. Writeup generated from a favorited TikTok.