agent sidequestsIdeas worth trying

Check AI piano feedback before changing how you practice

Use AI piano feedback as a set of questions to check against your recording. Before changing your playing, confirm the exact score passage and repeat, replay the cited moment, and separate what you can hear from what a tool inferred. The useful finish is one reviewed passage card with a small next experiment—not a score for your musicianship.

What can an agent usefully do with a practice take?

An agent can help organize a recording, your notes and the questions you want to revisit. Whether it can analyze sound is a separate capability to establish. A confident paragraph about a performance is not evidence that the assistant listened to it.

In his September 2026 Teach Me Again, AI account, José Luis Miralles describes useful annotation tooling alongside flawed musical feedback. The setup relied on tool-derived analysis rather than direct hearing, and he found reference and interpretation problems. That is one practitioner's experiment; we did not reproduce it.

The method below is our original, untested review process. Its worked example is fictional. It teaches you to decide what to correct in a report before deciding what to practice.

Bring a small reference packet

Start with one short passage and one purpose. Use material you may share, and keep the original recording. For Muse, Instinct, Dots or Grok Bot, check the tools available in your account; this guide does not establish compatibility.

  • The exact audio file and a way for you to replay it, with its filename visible.
  • The score excerpt you actually used, identifying its edition or arrangement, bar labels and any repeats or omitted bars. Resolve a mismatched reference before analysis.
  • One intention in your own words, such as making the last phrase feel deliberate, plus any relevant teacher notes.
  • A brief capability statement from the assistant: direct audio input, named analysis tool, or only organizing your supplied observations. If audio review is unavailable, use your timestamps and listening notes.

Check the location, the observation and the judgment

Ask the agent to turn each proposed correction into a checkable row: file, time range, score location and occurrence, what supports the claim, and what remains uncertain. Keep the original comment beside the correction so the same unsupported claim does not return later.

  1. Locate it. Play the cited interval yourself and confirm that it contains the intended passage. A repeated section needs an occurrence label, such as second pass. If you cannot establish the location, withhold the musical conclusion.
  2. Identify the evidence. Is the row based on your listening, a tool's estimated notes or timing, or an interpretation in prose? If the tool output and the recording disagree, do not silently treat the tool output as the ground truth.
  3. Test the meaning. Compare the proposed correction with your stated intention and teacher guidance. A difference from a reference is a question to investigate; it does not decide what you should want the phrase to sound like.
  4. Choose a disposition. Repair the reference, leave the claim unresolved, or keep a specific listening question. Select a practice experiment only after you can identify the relevant passage and explain what you want to try.

Worked example: a comment points to the wrong repeat

Fictional source packet: a player records a self-written eight-bar study as Take-A.wav. Bars 5–8 are played twice. The supplied score image shows that repeat. The player's goal is a deliberate ending on the second pass, without losing the sense of a continuous phrase. The player confirms first-pass bars 5–8 at 00:10–00:16 and second-pass bars 5–8 at 00:18–00:24 by listening.

Suppose an agent returns the three comments below. The decisions illustrate how to review a report; these are invented observations, not results from an audio test.

Three fictional comments with three different dispositions
Agent commentWhat the player checksReviewed decision
The second pass rushes at 00:12.00:12 belongs to the first pass. The agent attached the comment to the wrong occurrence.Correct the occurrence label and withdraw the second-pass claim. Relocating a timestamp does not prove the original judgment.
A bass note is missing at 00:20.The estimated note list omits it, but the player cannot resolve the sound in this recording.Keep it unresolved. Bring the excerpt to a teacher or make a clearer take; do not prescribe a fix for an unconfirmed missing note.
The ending is too slow.The player intended to linger. On replay, their own concern is a pause before that ending.Replace the blanket verdict with the player's question: can the ending feel deliberate while the phrase stays connected?

Leave with one reviewed passage card

Keep the next attempt small enough to describe in one sentence. A completed card should tell your future self where to listen, whose observation matters and what remains undecided. It should not turn an unresolved claim into a task by omission.

Illustrative completed card: Take-A.wav, 00:18–00:24; bars 5–8, second pass, in the supplied study. Location: confirmed by the player. Player observation: I notice a pause before my ending. Next experiment: play this passage again with my intended ending, listening for whether the lead-in feels connected. Open question: the reported missing bass note remains unresolved. Withdrawn claim: second-pass rushing at 00:12 was attached to the wrong occurrence.

Ask your agent to preserve that card in an editable note or in the replay page linked below. Keep rejected claims out of the practice queue, while retaining their correction history in the note.

Compare the next take using its own locations

In the fictional example, Take-B.wav reaches the same second-pass bars 5–8 at 00:23–00:30. Reusing Take A's seconds would play a different passage. Confirm both intervals before making an A/B listening pair, and label each file and occurrence visibly.

CAAMP's maintainer documentation describes correcting estimated score-to-audio alignment and working with partial mappings. Its stated limitations include omitted grace notes and timing-only annotations. We have not installed or tested this research software; it illustrates why a mapping is something to inspect, not a verdict on a performance.

Listen to A, then B, with the same question. Record your answer in plain language, including 'I cannot tell.' If you tried a change and prefer the result, keep that as your observation about these two takes. It does not establish durable improvement or explain every cause of the difference.

Before using an agent-built replay page, check that each button plays the named file and intended interval, Stop works, and edited boundaries survive reopening. If the player or analysis is unavailable, use an ordinary audio player and the passage card. You can still complete the review without automatic transcription.

Copy a prompt and try it

Each idea has a complete starter prompt, the inputs to bring, and the tools it needs.

Common questions

Can an AI assistant grade my piano playing from a recording?

This guide does not establish a reliable grade. Use the assistant to organize evidence and specific questions. Keep artistic intentions and unresolved disagreements visible, and use a teacher when you need instructional judgment. Do not infer physical technique from this review alone.

What if my assistant cannot hear the audio?

Have it work from your listening notes and confirmed timestamps. It can format the reviewed card or build playback controls without claiming to have assessed your performance. Label any tool-generated observations by their source.

Should I upload the whole recording archive?

Start with the smallest excerpt and reference needed for the question. Check where processing happens and what the service retains before uploading. Use recordings and scores you have permission to share, and do not include another person's lesson or playing without their permission.

What counts as a finished review?

You can replay the named passage, explain the evidence behind the surviving question, identify any withdrawn or unresolved claims, and choose one next attempt. A corrected report with no practice change can also be a useful finish.

Sources and further reading

These are practical editorial examples. The prompts have not been independently run-tested across every assistant.

Keep exploring

For platform setup, read the Muse, Instinct, Dots and Grok Bot guides.