/check-take

name: check-take
description: Checks a track's Suno takes against the track file by asking Suno's listening agent - which lyric lines are not sung or are changed, tempo, voice, excluded instruments, mix problems. Free and advisory; it never picks a take. Use after generating, or when the user asks "did it sing everything", "check the takes" or "is the tempo right".
argument-hint: <album-name> <track> [take N | all]
model: sonnet
effort: low
allowed-tools:
  - Read
  - bitwize-music-mcp
  - mcp__suno__*

Your Task

**Input**: $ARGUMENTS

Check a track's takes against its track file and report what differs. Suno's listening agent does the listening; you ask, collect and report.

You never choose a take and never rank takes. The report says what was heard; the user listens and decides.

---

Check Take Skill

Asking Suno's agent about a song is free: nothing is made, changed or charged. The limits are five unanswered questions at a time and 200 questions a day.

The agent is good, not infallible. Treat every flag as "listen here", never as proof. Background, and why Suno's lyric alignment is not used for this: `${CLAUDE_PLUGIN_ROOT}/reference/suno/mcp-api-mapping.md` § Take Check.

Step 1: Plan

Expected format: `<album-name> <track> [take N | all]`

Call `get_take_check_plan(album_slug, track_slug)`.

• `found: false` → report the error and stop.
• `takes` is empty → "No Suno takes to check for [track] yet — generate it first. (Takes logged as older share links cannot be checked.)" Stop.
• Which takes: `take N` → that one; `all` → every entry in `takes`; neither → `default_takes`.
• When `/bitwize-music:generate` handed off to you and `enabled` is false, stop without a word: the user turned the automatic check off (`generation.take_check: false`). A direct request always runs.

Step 2: Find the Tools

Look for `audio-listen` and `audio-listen-wait` on a connected MCP server and load their definitions. Not there → say the take check needs Suno's MCP server, give the setup command from the reference doc, and stop. There is no fallback.

Step 3: Ask

For each take, for each entry in `questions`, in order: `audio-listen(id: <clip_id>, question: <the question, word for word>)`. Keep the `id` each call returns, with its take and question number.

• Never have more than five questions unanswered. Two takes with one or two questions each fit in one round; beyond that, work in rounds of five.
• `429 concurrency_limit_exceeded` → read some answers first, then carry on. `429 quota_exceeded` → the day's 200 questions are used; say so and stop.

Then read each answer with `audio-listen-wait(id)`. It returns after about 50 seconds whether or not the answer is ready: call it again while `status` is `processing`. On `error`, note the reason and use an empty string as that answer.

Step 4: Evaluate

Once per take: `evaluate_take_check(album_slug, track_slug, clip_id, answers)` — that take's answers word for word, one per question, in question order. Pass `duration_s` when you know the take's length.

Step 5: Report

TAKE CHECK — [NN] [Title]

Take 1 (row 3) — 2 things to check
  ⚠ Lyrics   line 2 sung with different words: "Paper boats are lined up on the sill"
             heard: 'all lined up' instead of 'are lined up'
  ⚠ Tempo    112 BPM measured, 96 asked
  ✓ Vocal    female lead — soft, breathy

Take 2 (row 4) — nothing flagged
  ✓ Lyrics   all 24 lines sung as written
  ✓ Tempo    95 BPM (asked 96)

These are the agent's observations — listen at the flagged spots before deciding.

• `warn` checks first, then `ok`. `info` checks (the key, an unread answer) go last, and only when they add something.
• A few lines per take. Quote each check's `detail`; do not turn it into a verdict of your own.
• No recommendation, no "take 2 is better", no ordering by quality.

After the Report

• The user rejects a take over a flagged problem → the Regeneration Workflow in CLAUDE.md: a lyric or pronunciation problem goes to `/bitwize-music:lyric-writer` and `/bitwize-music:pronunciation-specialist`, a style problem to `/bitwize-music:suno-engineer`.
• A flag looks wrong to the user → the agent can mishear, and its answers vary a little from run to run. Offer to ask again: `/bitwize-music:check-take <album> <track> take N`.
• Nothing is written to the track file. The Generation Log stays the user's record.

Error Handling

**Suno's server is not connected:**

The take check asks Suno's listening agent, so it needs Suno's MCP server:
  claude mcp add --transport http suno https://api.suno.com/mcp
then run /mcp and sign in.

**An answer has no readable result** (the report's `unread` check): say which lines were not checked and offer to ask again.

Remember

1. **Advisory only.** Report what was heard; never pick, rank or reject a take. 2. **Word for word.** Questions go to Suno exactly as the plan gives them; answers go to `evaluate_take_check` exactly as Suno gave them. 3. **Five questions at a time, 200 a day** — and it is free. 4. **A flag is a place to listen**, not a verdict: the agent can be wrong.