An effect written by AI, checked by the AI itself: “the harder I play, the longer it repeats”

The actual test conversation took place in Japanese. This log is a direct English translation of it.

User

I want an effect where the harder I play, the longer the notes repeat.

AI (Claude Code)

I wrote “Velocity Repeat” and ran it through MIDIFabric’s Check. It passed with no errors or warnings. I have not submitted it yet.

One thing to tell you first about this version. In this conversation I already know the problems of the earlier version, such as repeats of the same note cutting each other off and the later repeats becoming inaudible. Rewriting the same mistakes while pretending not to know would make the demo staged, so I designed it with the fixes from the start. What this demo can verify is the flow: writing → Check → registration → checking by reading the routing and the monitor. It does not show what an AI produces on its first try; that needs a separate session started in an empty folder.

Behavior

• Count: proportional to the velocity you play, calculated as round(velocity / 127 × Max Repeats), at least once.
• Level: the repeats fade evenly so that the last one plays at “played velocity × Tail” (default 0.35). Even when you play hard, every repeat stays clearly audible.
• Playing the same note again: stops the earlier repeats, cancels the Note Off of the sounding repeat and stops it right away.
• When a repeat starts while the key is still held: the repeat takes over that note, and the Note Off from releasing the key later is dropped.
• Control panel: Interval (knob, default 150 ms), Max Repeats (8), Tail (0.35), Gate (0.5).
• MIDI 2.0 notes: passed through unchanged.

May I submit it? With automatic registration on, it will be registered as soon as it is submitted. Once it is registered, please place it on a route. After that, I will read the routing to check where it is placed and its settings.

User

Please submit it.

AI (Claude Code)

I submitted it, and it was registered automatically.

• Submission: MIDIFabric received it. Check reported nothing, and the candidate revision is 1.
• Registration: I read it back over MCP to confirm. “Velocity Repeat” is in the Library (ID 1E05E02D-…), and the registered Lua code is identical to what I submitted.

Next, please:

1. Place it: on the routing canvas, put “Velocity Repeat” between a keyboard and a synth and press Apply. Any keyboard and synth will do. To match last time, connect the KeyStep Pro or the Prophet Rev2 to the built-in synth or the JUNO-D8.

2. My check: tell me once it is placed and applied. I will read the routing and check the placement, connections and settings.

3. Try it: after that, clear the monitor once and play these three:

   • once softly
   • once hard
   • the same note hard two or three times in a row

   When you have played, just tell me “played it”. I will read the monitor records and count.

User

Placed it, pressed Apply and played. Check the MIDI log.

AI (Claude Code)

I read the routing and the monitor over MCP and checked. There were 222 records in total, and I read all of them in two pages. Nothing was dropped, the monitor was not paused, and no record was truncated.

Routing

• KeyStep Pro → Velocity Repeat → Midi Out 2 (the JUNO-D8 through the MIDI4x4) is connected and enabled in Route 1.
• The settings were at their defaults (Interval 150 ms, Max Repeats 8, Tail 0.35, Gate 0.5).

Repeat counts and levels: all as calculated.

• The last repeat is always “played velocity × 0.35” (44 for v127). Unlike last time, no repeat is too quiet to hear.
• The interval is about 150 ms and each repeat lasts about 75 ms (Gate 0.5), consistently.

Playing the same note again (twice)

• When G3 was played again during its repeats, the Note Off of the sounding repeat came just before the new note (within 0.4 ms), and then the new note started.
• The rest of the old repeats and their reserved Note Offs did not appear. No new note was cut off by an old Note Off, and every repeat sounded for about 75 ms. The problem from last time, where notes were cut after 0.01–5 ms, is fixed.

Different notes overlapping: the repeats of C3 and E3 ran in parallel without interfering with each other.

Other: the channel pressure (D0) sent while playing passed through unchanged.

Not yet verified

• A repeat starting while a key is held: this time every key was released before the first repeat (at 150 ms), after 80–116 ms. The takeover of a held note has not run yet. To check it, play one note and hold it for about half a second.
• How it sounded: the monitor only shows the MIDI that was sent. Please tell me how it sounded on the JUNO-D8 (did harder notes feel like they repeated longer?).

User

Played while holding the keys. Check the log.

AI (Claude Code)

Holding the keys also worked as intended. I read all 146 new records. Nothing was dropped and the monitor was not paused.

Four notes where the repeat started while the key was held: all behaved the same way. At the first repeat, the original note is stopped (Note Off) and then the repeat sounds. The Note Off from releasing the key later was not sent out.

• Note Off on release: all four were recorded in the monitor as “blocked” (Velocity Repeat received it and did not pass it on). The records show directly that the effect stopped them on purpose.
• Were any repeats cut off? I played E3, F3 and G3 while B3’s repeats were still going, but every repeat sounded for about 75 ms and none cut another off.
• For comparison: the last note, F3 v112, was released before its first repeat. In that case the Note Off on release went out as is, the same as in the earlier check.

Other things in the records MIDI 2.0 notes from the A-88MK2 passed through Arpeggiator1 and Echo1. That path is turned off in the routing, so nothing went out to a device. It is unrelated to Velocity Repeat.

What has been verified so far

• The count changes with how hard you play
• Every repeat stays loud enough to hear, to the last one
• Playing the same note again stops the earlier repeats
• Repeats of different notes do not interfere with each other
• When a repeat starts while the key is held, the repeat takes over the note and the Note Off on release is stopped

All of this was verified as MIDI output.

What remains is how it sounded. The monitor only shows the MIDI that was sent; I cannot tell how it sounded on the JUNO-D8. Please tell me whether harder notes felt like they repeated longer, and whether anything bothered you.

User

Playing hard made it last longer, and I could hear it properly.

User

Make it a bit shorter when I play softly, and let me adjust the pitch change with a knob too.

AI (Claude Code)

I updated Velocity Repeat for the second request, “Make it a bit shorter when I play softly, and let me adjust the pitch change with a knob too.” It passed MIDIFabric’s Check with no errors or warnings. I have not submitted it yet.

What I changed

1. Shorter when played softly: I added a Velocity Curve slider (1 to 3, default 2) and changed the count to round((velocity / 127) ^ Curve × Max Repeats). With the curve applied, hard notes still repeat up to the maximum as before, while soft and medium notes repeat fewer times.

   Velocity played	Count now	After the change (Curve 2)
   29	2	1
   50	3	1
   76	5	3
   94	6	4
   114	7	6
   127	8	8

   Setting Curve to 1 gives the same counts as now. You can adjust how much “shorter” with this slider.

2. Pitch change with a knob: I added a Pitch Step knob (−12 to +12 semitones, default 0). Each repeat moves up or down by that many semitones. For example, with +2, playing C3 goes up D3, E3, F#3 and so on. When a repeat would go beyond the MIDI note range (0–127), the repeats stop there. At the default of 0, it repeats the same note as before.

Panel layout I kept the layout you arranged in the app (style “real” and the position of each control). The Pitch Step knob is under the Interval knob, and the Velocity Curve slider is under Tail.

A caveat When Pitch Step shifts the pitch, a repeat can land on the same note as another key you are holding. In that case, the repeat’s Note Off may stop the note you are holding.

May I submit it as an update to the existing Velocity Repeat? With automatic registration it should be updated right away. Once it is registered, I will check the settings in the routing. After you play, I will count the repeats and pitches in the monitor.

User

Please submit it.

AI (Claude Code)

I submitted the update and confirmed, by reading the routing, that it was applied automatically.

• Submission: MIDIFabric received it. The candidate revision is 2, and Check reported nothing. There was no editing_conflict this time.
• Applied: the definition revision of the Velocity Repeat on the route went up from 1 to 2. Two new settings were added:
  • Velocity Curve 2
  • Pitch Step 0
• Unchanged: the placement (KeyStep Pro → Velocity Repeat → JUNO-D8) and the other settings (Interval 150 ms, Max Repeats 8, Tail 0.35, Gate 0.5).

Next, please:

1. Set Pitch Step: turn the Pitch Step knob on the control panel to, say, +2 (each repeat goes up a whole tone). At 0, the pitch change cannot be checked.
2. Clear the monitor and play these three:
   • once softly (velocity around 50)
   • once at medium strength (around 80)
   • once hard
3. Let me know: when you have played, just tell me “played it”. I will check the Pitch Step value in the routing and count the repeats and pitches in the monitor. I will also check whether soft notes got shorter compared with the previous version.

User

Played it. Check the log.

AI (Claude Code)

I read the routing and the monitor to check. “Shorter when played softly” works as intended. However, the pitch change could not be checked from this playing.

The monitor had not been cleared, so I read only the 118 records made after the previous check. Nothing was dropped.

Counts when played softly (Velocity Curve 2)

• Soft and medium notes repeat fewer times, while hard notes still repeat 8 times. The last repeat is always about 0.35 times the played velocity.
• The first B3 was played again after 3 repeats, so a new chain restarted with 8 repeats.
• The takeover when B3 was held, both times, also worked correctly. The Note Off on release was recorded as “blocked”.
• Overlapping repeats of B3 and C4 did not interfere with each other.

The pitch change has not been verified yet.

• Reading the routing, Pitch Step was still 0.
• In the monitor, too, every repeat has the same pitch as the note played.

If you did not turn the Pitch Step knob on the panel, this is the expected behavior.

Please tell me whether you turned the knob.

• If you did not: set Pitch Step to +2 or so and play again. I will count how the pitch rises in the monitor.
• If you did: the value has not reached the routing. I will look into it as a problem either in the panel or in reading the setting.

User

Set Pitch Step to +2 and played. Check the log.

AI (Claude Code)

With Pitch Step +2, the pitch change also worked as intended. Reading the routing, Pitch Step is 2, and I read all 120 new records in the monitor (nothing dropped).

• Pitch: each repeat rises by a whole tone (2 semitones).
• Note Off: each one went to the shifted pitch. No note was left hanging.
• Count and level: as calculated with Velocity Curve 2. Soft notes ended after one repeat.
• Overlap: playing B3 and C4 hard one after the other, the two rising chains overlapped without cutting each other off.

So the second request from the product page, “Make it a bit shorter when I play softly, and let me adjust the pitch change with a knob too”, has been verified in the MIDI output. If you like, please tell me how it sounded too.

If you want to continue, there is the third example on the product page: “How does this Lua decide the number of repeats from how hard I play?” That is a question, so I will leave the effect unchanged and explain it by reading the registered code.

User

How does this Lua decide the number of repeats from how hard I play?

AI (Claude Code)

The count is decided once, at the moment the key is pressed, by a function in the registered code called repeat_count.

local function repeat_count(velocity, max_repeats, curve)
  local count = math.floor((velocity / 127) ^ curve * max_repeats + 0.5)
  return math.max(1, math.min(max_repeats, count))
end

The calculation has four steps.

1. Turn the strength into 0–1: divide the velocity by 127. v127 becomes 1.0, and v64 about 0.5.
2. Apply the curve: raise that value to the power of Velocity Curve. With the default of 2, it is squared. Values below 1 get smaller when raised to a power, and softer notes shrink more. 1.0 (the hardest) stays 1.0 whatever the power.
3. Multiply by the maximum and round: multiply by Max Repeats, add 0.5 and drop the fraction. In other words, round to the nearest whole number.
4. Keep it in range: keep the result between 1 and Max Repeats. However softly you play, there is at least one repeat.

With the current settings (Max Repeats 8, Velocity Curve 2):

The counts read from the monitor match this table (1 for v29, 3 for v72, 5 for v99 and 8 for v127).

Notes

• When it is decided: the count is fixed when the key is pressed and stored when that note’s repeats are created. Moving a slider during the repeats does not change the count of the repeats already playing; it applies from the next note you play.
• The level is calculated separately: the velocity of each repeat is set by Tail ^ (repeat number / count), independently of the count. Whatever the count, it fades evenly so that the last repeat is “played velocity × Tail”.
• With Velocity Curve at 1: it becomes simply proportional (velocity / 127 × Max Repeats), the same counts as the version from the first request.