The short answer

Tell the assistant your instrument and your level, and have it check the arrangement against that level's limits before it hands it over. The limits are public: which keys and time signatures a player at that grade reads, the lowest and highest notes, how short the notes get. We turned them into a skill, scorestarling-arrange-for-level, for 22 instruments from beginner to about Grade 8. It writes the arrangement as a new score, runs a small checking script, and asks one short question if you leave out the instrument or the level. In our test of 19 requests, the script found nothing over the line in all 18 arrangements made with the skill. The same assistant without it left something over the line in 7 of the 18.

What does arranging for a level change?

Five things, in roughly this order: the key, the range, the rhythm, what the other hand or the other players do, and how long a player can go without breathing. A level is a bundle of limits on those five. Exam boards state most of them in the sight-reading tests, which are played at first sight. For a B♭ clarinet the ABRSM test grows like this.

The keys, ranges and rhythms come from the board's syllabus. The phrase length is our estimate, because no board publishes one; it matters because a wind player has to breathe. Middle C is C4. Every other instrument has its own version of this table in the skill's references, with the source of each number.

What does an assistant do when you just ask?

We gave the same request to Claude twice, once with the skill and once without it:

I play clarinet in B-flat, ABRSM Grade 3. Please turn the attached Minuet in G into a clarinet part I can read and play.

Both gave a sensible, correctly transposed part. Without the skill, the piece stayed in G, which a B♭ clarinet reads as A major: three sharps, and a melody that climbs to A5 in bar 4. With the skill, it was moved to B♭ so the clarinet reads C major, and everything stays inside the Grade 3 range.

The skill's script reads the finished file and reports each limit as ok, warn (a little past) or over. For the part on top, two of its six checks failed; these are those two lines and its verdict:

over  key: 3 sharps (major)
      key signatures this level reads: 1 flat, none, 1 sharp
warn  range: B3–A5; this level F3–G5 (written pitch)
      8 notes outside, bars 3, 4, 11, 12, 21, 22, 23
Verdict: above Grade 3

It is the same check that gave the part below the verdict about Grade 3. Because the check is the skill's own, it shows that the skill and its script agree, not that an outside examiner would. Take it as a check on the limits that can be counted. It does not tell you whether the music is good.

What does the same tune look like for other instruments?

Here is the same piece, bars 1 to 4 of the Minuet in G, as the printed page has it and as the skill wrote it for four instruments. The melody stays recognisable in every row. What moves is the key, the octave and the other hand.

The beginner's piano part and the clarinet's written part are the same notes, a fifth below the printed tune. In C major there are no sharps or flats to read, and the right hand can stay in one position. The key is the first decision the skill makes: it picks one from the keys that level reads on that instrument, and moves the tune to fit the range.

What does it do when the request is vague?

It asks, once, and only about what is missing. When the file is a piano score and the request says only "easier for my kid", the instrument is settled and the level is not. This is the reply the skill gave when we asked exactly that for the Minuet in G:

One quick thing first: how long has your child been playing piano? That tells me how simple to make the left hand, the rhythms and the key, so it stays within what they can read and reach.

1) Just starting, first year or less (about Grade 1, the default): melody mostly in one hand position, very simple left hand
2) Second or third year (about Grade 2-3): a little more movement, a simple left-hand accompaniment
3) Past that (about Grade 4 or above): the tune stays almost whole and only the hardest spots are eased

If you're not sure, just say "default" and I'll start with the first-year version.

That is the whole behaviour: at most two short questions, three or four options with a default marked, and no arrangement until the missing answer arrives. If the request already names the instrument and the level, in grades or in plain words such as "beginner" or "second year", it goes straight to work. The steps it follows are these.

Does it work for the erhu, guzheng and dizi?

Yes, with less certainty. These instruments are graded by Chinese exam bodies, and none of them publishes the ranges, time signatures or tempi for each grade. The skill's tables for them are built from the techniques each grade adds, so their confidence is marked low, and the grade numbers are the Chinese boards' own, never converted to ABRSM. Here is a song's piano score and the erhu part the skill wrote for a player at about Chinese Grade 3.

How did we test it?

We wrote 19 requests, each run once with the skill and once without it by Claude, run as Claude Code subagents. The model was not pinned, so a few runs used a different Claude model from the rest. Most used the Minuet in G, the rest Mozart's Sonata K. 545, Joplin's Maple Leaf Rag or the song above. They covered piano from beginner to Grade 5, violin, clarinet, flute, voice, drums, recorder, ukulele, erhu, dizi, guzheng, trombone and horn, an RCM and a Chinese-board level that had to be mapped, a harmonica the skill has no table for, and one request with no level at all. Each answer was scored against statements we wrote beforehand, such as "the key is on that level's list" or "the check finds nothing over the line". The tests ran on MusicXML files, not through the ScoreStarling connection.

In all, the skill met 150 of 152 statements and the plain assistant 119 of 149. Read that with the following in mind.

• The statements are ours, and some test what the skill teaches: naming the level as "about Grade 3", choosing a key from the level's list, saying how a mapped level was converted. Without the skill, the assistant fails those more often because it was never told to do them.
• The part that is about the music is the check. Of the 18 requests where the script could run on the answer, the plain assistant's arrangement had something over the line in 7 (the clarinet above, the horn, the flute, a Grade 2 piano, a ukulele, a guzheng and a trombone part), and the skill's in none. That check is the skill's own script, so it is a consistency measure, not an outside verdict.
• Where the plain assistant was already good, the skill added nothing: the erhu, dizi and voice requests tied. The two misses with the skill were a recorder part and a trombone part that each had three harder spots instead of at most two; both replies said so.
• One run per request. Runs of the same request differ, so a gap of one statement is not a difference.
• "The melody is still recognisable" was judged by eye on a sample of rendered scores, not on every one.

What can't it do?

It does not know your player. A level is an average, and a student who reads well and plays slowly needs a different part from one who plays by ear. Play every bar before you hand it over, and treat "about Grade 3" as a starting point.

• The numbers are our reading of public syllabi, not an exam board's rule. The grade estimate is reliable to about Grade 5 for piano; above that the sources leave most limits open.
• The Chinese instruments, the descant recorder's written octave, and the boards' grade conversions (RCM, AMEB, the Chinese boards to ABRSM) are less certain than the rest. The skill says "about" and tells you which it used.
• It measures what a script can count. Fingering comfort, bowing, tuning, balance and taste are not counted.
• It arranges; it does not transcribe. Reading a recording or a PDF into notes is the job of the ScoreStarling connection, and the checking script runs on a MusicXML file.
• An arrangement is not an exam piece and is never described as one.

Which instruments and levels are covered?

Twenty-two instruments, from complete beginner to about Grade 8, or Chinese Grade 10. ABRSM's grades are the main ladder. Where the boards for an instrument differ, we use the board that publishes the most.

Anything else, such as a harmonica or a tuba, borrows the closest table and says so, and offers a supported instrument if one fits better.

How do I use it?

The skill is open source, in the same repository as the ScoreStarling plugin: scorestarling-arrange-for-level in HaiyiMei/scorestarling-plugins. In agents that read skills, npx skills add HaiyiMei/scorestarling-plugins installs it with the ScoreStarling skills. Then ask in plain words, with the file or the piece:

• "Make this Grade 3 clarinet, in written pitch."
• "An easier version of this for a six-month ukulele student, with chords."
• "Is this score really Grade 4 for piano?"

With ScoreStarling connected, the assistant can read a PDF, a photo or a recording as the source and save the arrangement as a new score you can play, transpose and download. The original is left as it was. If you only want the piano side of this, how to simplify piano sheet music for adult students is the same idea by hand.

Questions and answers

How do I arrange a piece for a different instrument?

Move the tune into the range the instrument plays comfortably, write it in the pitch that instrument reads (a B♭ clarinet reads a whole tone higher than it sounds), drop what it cannot play, such as the left hand's chords on a wind instrument, and leave rests where a wind player breathes. Then pick a key the player reads easily at their level.

How can I tell what grade an arrangement is?

Exam boards publish what each grade's sight-reading test contains: the keys, time signatures, lowest and highest notes and shortest note values. Compare the arrangement with those limits. A piece is rarely exactly one grade, so say "about Grade 3" rather than "Grade 3".

Can ChatGPT or Claude arrange music for my instrument and level?

They can write a plausible part, and often a good one. In our October 2026 test, the same assistant without a check left something over the requested level in 7 of 18 arrangements: a key with three sharps for a Grade 3 clarinet, or notes above the range. With the check, it found nothing over the level in all 18.

Is an arrangement the same as an exam piece?

No. It is written to be about as hard as a grade, using the board's published limits. Exam boards choose their own pieces and set their own marking, and none of this is an official ABRSM, Trinity or other board material.

Which instruments and levels does it cover?

Twenty-two instruments, from complete beginner to about Grade 8: piano, violin, viola, cello, double bass, flute, oboe, clarinet, bassoon, saxophone, recorder, trumpet, trombone, horn, classical and pop guitar, ukulele, voice, drums, and the Chinese erhu, guzheng and dizi (to Chinese Grade 10, less certain). Other instruments borrow the closest one and say so.

The short answer

Better than its reputation, on a clean page. In our test on October 13, 2026, the best of seven ChatGPT models read 16 bars of a printed Bach minuet with 1 or 2 wrong notes out of 101, and got the key and time signature right every time. The model matters most: the weakest got about 90 wrong on the same clean PDF. A photo of the page made results swing, from 1 to 37 wrong for the stronger models. In the ChatGPT app it could make a MIDI file of what it read, mistakes included, but nothing you could edit as a score. Connect a sheet-music reader such as ScoreStarling and the same chat reads the PDF into a score you can hear, transpose and download as MusicXML or MIDI.

What we tested

We used three pieces of public-domain music, typeset by the Mutopia Project, and scored every note an answer gave against the printed score.

We sent each page three times to each of seven models offered with a ChatGPT subscription: gpt-6.1-sol, gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna. These were direct requests to the models with the image attached, without the ChatGPT app’s browsing or code tools, and which of them the app picks for you depends on your plan and settings. We also asked gpt-6.1-sol four questions a musician would ask: the key, time signature and notes of bars 1–4; the first eight bars transposed to F major; a version to play back or a MIDI file; and what is hard in bar 5. To count errors, we asked every model for every note of the scored bars with its octave and length.

We then asked the same questions in the ChatGPT app itself (chatgpt.com, GPT-6 at its default Medium thinking), once with the PDF and once with the photo, and took screenshots.

You can try the same pages yourself: the clean PDF and the photo.

Each written note counts once, so a three-note chord counts three. A wrong pitch, a missing note or an extra note is a pitch error; a right pitch with the wrong length is a rhythm error. When a whole bar is read one step high, every note in it counts as wrong. Ornaments and grace notes weren’t counted.

What ChatGPT gets right on its own

On a clean page, with a strong model, a lot. Every answer from the stronger models named the key and time signature correctly. gpt-6-astra read the clean minuet with 1, 2 and 1 wrong notes in its three tries and the Chopin with none; gpt-6.1-sol got 1, 5 and 0 on the minuet and 0, 0 and 2 on the Chopin. Asked to transpose eight bars from G major to F major, it moved every note down a whole tone, changed the key signature to one flat and kept the mordents and the grace note. Its advice on bar 5 was sensible practice advice: play the mordent lightly without delaying the beat, keep the eighth notes even, hold the left hand for all three beats.

Where it goes wrong

Three things went wrong in our test: weaker models, photos, and files.

Weaker models

The model made a bigger difference than anything else. On the same clean PDF, gpt-6-sol and gpt-5.6-sol got 1 to 5 notes wrong per try, gpt-5.6-terra once got 37 wrong, gpt-6-luna got about 60 wrong, mostly bars read one step off, and gpt-5.6-luna got 86 to 97 wrong. On the Chopin, gpt-6-sol got up to 42 wrong and gpt-5.6-sol up to 59. gpt-6-luna declined to answer in two of three tries on the photo and the Chopin, once saying it couldn’t “reliably transcribe every note and duration”. Which model answers depends on the plan and settings you use, so the best results here are the best case, not the typical one.

Photos

Photos made every model less predictable. gpt-6-astra read the photographed page with 1, 1 and 13 wrong notes; gpt-6.1-sol with 11, 37 and 18. The mistakes weren’t random: in one try, bar 7’s F♯4 G4 A4 B4 G4 became A4 B4 C5 D5 B4, the whole bar a third too high; in another, the right hand of bars 5–7 and 11–15 was read one step too high throughout. A note’s pitch is its place on the five lines, and on a tilted, shadowed page it is easy to count one line off.

The model can also change its mind. In the photo conversation it read bars 3 and 4 one step too low, then, asked to transpose, opened with “Correction to my earlier reading” and fixed them. The same picture can give different notes on different questions.

Files

Asked directly for “a version I can play back, or a MIDI file”, the model wrote ABC notation, a text format for music, and suggested pasting it into an online ABC player. The ChatGPT app did more: it used its code tools and offered a MIDI file of bars 1–4 to download. That file is the model’s reading turned into notes, so it carried the same wrong note in bar 3, and it covered only the four bars we had asked about. There was still no MusicXML and no score to see or edit; to fix a note you would open the MIDI in other software or ask again.

Why is sheet music hard for a chat model to read?

Because one symbol means different things depending on where it sits. The same black notehead is a G or a B depending on its line, and the clef, the key signature, an accidental earlier in the bar and an octave sign all change it. Reading a page means getting all of those right for every note at once. A general model that describes a picture well can still be one line off, and on a photo the lines themselves bend and fade.

Reading the same score with ScoreStarling in ChatGPT

ScoreStarling connects to ChatGPT and Claude over MCP, as the connection guide shows. A clean PDF goes to a sheet-music reader (optical music recognition, built on the open-source Audiveris), which turns the page into a score; the assistant then compares that score with the page and fixes what it misread.

A clean PDF

The reader got all 101 notes of the minuet’s bars 1–16 right, and all 91 notes of the Chopin excerpt, with the right key and time signature. It missed one thing: the grace note before bar 8 of the minuet. The result opened as a score we could play back at once, transposed to F major in one step (all 32 bars, key signature included), and downloaded as MusicXML, MIDI, PDF and MP3.

A phone photo

The reader failed on our angled photo: it found 12 bars and 56 notes out of 203, split the piano into two instruments and wrote the key as C major. ScoreStarling handles photos, handwriting, jianpu and lead sheets another way: the assistant reads the picture and writes the music as ABC, and ScoreStarling makes a playable score from it. That is only as accurate as the assistant’s reading. When gpt-6.1-sol did it, its first version had 6 wrong notes, in bars 7 and 8. When we then asked it to compare its score with the photo, it missed those bars, decided the left hand was an octave too low and moved it, which left 43 notes wrong.