Claude Code music generation, reviewed: the bed for a 30-second demo video

Attempt 1: a sentence, and the model sang it

The clip has a voiceover on top, so the music had to sit under a voice: no vocals, light drums, steady tempo. We wrote that as one sentence and did not name a model — the point was to see what the agent offered. Thirty seconds later two tracks came back, 10 and 12 seconds long. They had lyrics. The lyrics were our brief.

We did not trust our ears alone, so we ran the track through the transcription service for $0.10 per run. It came back with "Bright. … for a 30-second product demo. Light drums, warm keys, no vocals." — the model had treated the sentence as lyrics and sung "no vocals". A second attempt with the instrumental switch on but the brief still written as one sentence did the same thing at 36 and 39 seconds. The fix was not a better sentence; it was using the call's fields properly.

Attempt 3: the brief that worked

That is a real Claude Code session, unedited. Both candidates came back at about 95 seconds, and the transcription check on each returned no words. That is what an instrumental should give, but an empty transcript is weak evidence on its own — the service also returns nothing when it cannot follow the audio — so we listened to both before keeping one. For comparison we ran the same brief through MiniMax Music with the instrumental switch on; it returned one track of 2 minutes 42 seconds, also with no words, in just over three minutes.

What we would tell the agent next time

Say "instrumental" as an instruction, not as a word in the description — the agent then sets the switch on the call instead of hoping the model reads the sentence that way. Put the style where the style goes: Suno's call has a tags field for genre and mood and a negative tags field for what to avoid, and once the style lived there instead of in the lyrics slot, the singing stopped. Say the length as a number if the cut is fixed; "for a 30-second demo" produced 10-second tracks on the first run and 95-second tracks on the third, so trimming or looping is part of the job either way, and the agent does that locally for free.

How the session decides what to run

Before replying, the agent read the capability index the Skill keeps on disk, which is why its first message already listed three models with a price beside each — presented neutrally, no default, no recommendation. Say "you pick" and it takes the best-quality tier. Picking a model is the approval: the run starts at the shown price with no second question, and if the wallet cannot cover it the agent stops before anything is charged and gives you a top-up link with the amount prefilled. Every run went through Actionway's music and sound capability.

Music is priced per run, and a run can return more than one candidate — the Suno runs returned two each. Suno v5 and v5.5 both cost $0.20 per run; MiniMax Music is $0.40 per run. Four runs to get it right cost less than a coffee, and the voiceover and the subtitles for the same video came out of the same session. There is no subscription; one balance covers every capability in the catalog.

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 and an Actionway account; after sign-in the agent tells you your wallet balance and, if it is empty, links you straight to top-up. Prefer a plugin or a Remote MCP server to the command line? Both end at the same catalog and the same wallet — the get started page lists the steps per client side by side.