AceStepClaude_Lyrics

If you've got an Anthropic API key, this is arguably the best-sounding writer in the pack. AceStepClaude_Lyrics is structurally identical to its OpenAI sibling - describe a song, get back ACE-Step-formatted lyrics with production tags in [brackets] - but Claude's strengths (instruction-following, avoiding repetition, staying in format) line up suspiciously well with what the shared prompt template demands. People reach for this one when the other providers keep producing clichéd or mis-formatted lyrics.

The mechanism

Same skeleton as every lyrics node in the pack: the text field is your song description, the node wraps it in the shared lyricist prompt (short singable lines, one tag per line, blank-line separation, vocabulary-diversity rules), and sends it to api.anthropic.com/v1/messages with your x-api-key header. No SDKs - plain urllib, and the api_key is a paste-into-the-node input, not an env var. max_tokens defaults to 1024, seed goes into the prompt text to vary results, and instrumentality_balance steers how vocal-heavy the output is. The output is a single lyrics STRING that feeds your ACE-Step conditioning.

Which model

The dropdown defaults to claude-opus-4.5 - the flagship, and the one the author recommends. That's a genuinely good default for a creative writing task, but it's also the most expensive option in the family. If you're iterating on styles, claude-sonnet-4.5 gets you 90% of the way at a fraction of the cost, and claude-haiku-4.5 is the cheap-and-cheerful pick for bulk experiments. The snapshot entries (claude-3-5-sonnet-20241022, etc.) exist for people with locked-in API habits; you can ignore them.

Notes and gotchas

• The node sends no temperature parameter - Anthropic's API defaults apply. Fine in practice; don't go hunting for a temperature knob that isn't there.
• Errors surface as the lyrics string (e.g. Error: ...), same as the rest of the family. If the sampler gets weird text, look at the node output first.
• No retry, no rate-limit handling - a 429 comes straight back at you.
• API keys live in the node, so they end up in saved workflow JSON. Treat shared workflow files like a secret if you've pasted a key into one.

Install is the pack standard: ComfyUI Manager (search "JK AceStep Nodes"), or

cd ComfyUI/custom_nodes
git clone https://github.com/jeankassio/JK-AceStep-Nodes.git

then restart. The pack adds librosa on top of torch/numpy/tqdm; this node itself pulls in nothing extra.

The honest take: OpenAI's node is the default, Groq's is the free/fast one, and this is the one you reach for when lyric quality is the bottleneck - which, for vocal-forward ACE-Step tracks, it usually is.