Best Claude Code TikTok skill: from a brief to a clip

The finished clip

Start with the output, because the point of this is that a complete short video comes out the other end rather than a folder of assets. Two shots, one voice line, joined and mixed in one session.

Two shots, not one long take

The instinct is to ask for the whole video in one request. Do not. Video is priced per second of output, so a ten-second take costs the same as two five-second takes — but a single take arrives as one unbroken shot you cannot recut, and if one second of it is wrong the whole thing is wrong. Two shots give you a cut, an order you can change, and the option to re-run only the half that failed.

That last row is the one worth internalising. Concatenating two clips and muxing an audio track is a single command your agent runs on your machine. It does not go through a service, it is not billed, and it is not slower than the alternative. Any workflow that treats editing as something to buy is paying for the wrong half.

The resolution nobody warns you about

So the resolution decision belongs at the start, not the end. Draft at 480p while you are still deciding whether the shot list works — at these prices you can throw several away — then re-render only the shots that survived at a higher tier before you publish. That is the same two-pass approach the trailer guide uses, and it exists because the cheap pass is for judging structure and the expensive pass is for the thing people actually watch.

The voiceover is shorter than the footage

Our line ran 6.6 seconds against 10.2 seconds of picture. That mismatch is normal and it is a decision, not a defect: either the footage carries on after the voice stops, or you trim the picture down to the voice, or you write a longer line. We left the picture running, because a beat of silence at the end of a short video is how you let the last image land.

What you should not do is generate the voice first and then try to make the shots fit it to the frame. Speech length is not something you can specify — it comes out however long the model reads it. Generate the picture to the structure you want, generate the line, and settle the difference in the free local step where changing your mind costs nothing.

What to run and in what order

1. Write the shot list first

   Two to four shots, each doing one thing, each describable in a sentence. This is free and it is the part that decides whether the video works.

2. Draft every shot at the cheapest tier

   Watch them in order before spending anything on quality. Half the shots will change, and at draft prices that costs almost nothing.

3. Re-render the keepers higher

   Only the shots that survived, and only then. This is where the resolution gap gets closed.

4. Generate the voice last

   Once the picture is locked you know how much room the line has. Then join, mix, and trim locally at no charge.

Running it from Claude Code

Claude Code cannot render a frame of video or speak a line, but it can run ffmpeg, and that split is the whole shape of this workflow. The Skill installed with the CLI adds video generation and voice generation; the cutting is work the agent already knows how to do. Every run is priced before it starts, and an empty wallet stops the job before anything is charged.

Set it up once

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.claude/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.claude/skills/actionway"

You need Node.js 20 or newer, an Actionway account, and ffmpeg on your machine for the local half. Codex users copy the same folder into ~/.codex/skills/actionway instead, and both routes reach the same catalog and the same wallet — the get started page lists the steps per client side by side.