Roblox Sfx

Roblox sound, by id, verified

First time on this machine? Read LOCAL.md in this folder. If it does not exist, run the roblox-skills-setup skill (or copy LOCAL.template.md to LOCAL.md and fill it in). LOCAL.md holds your paths and tool locations; it is yours and is never shared.

The whole skill rests on one fact, measured not assumed:

A free Creator Store audio asset plays in an unpublished, unsaved place, with no API key, no ownership and no upload.

So sound does not need the Open Cloud audio quota (10 uploads/month unverified, 100 ID-verified, create-only, so every regeneration burns a slot). It needs a search and a check. That takes the audio stage off the critical path entirely.

Copy code
spec.json  (roles: what the game needs a sound FOR)
  -> find_sfx.py     unauthenticated toolbox search -> candidates.json
  -> emit_probe.py   probe.lua + report.lua
  -> Studio          create the Sounds, then MEASURE them   <- the gate
  -> pick.py         three gates, then score -> manifest.json
  -> emit_bank.py    SoundBank.luau + Sfx.luau + install.lua + tour.lua

Search produces candidates. Studio produces facts. Only facts choose.

What was measured, and when

Verified live against Studio on 2026-08-30 in an unsaved Place1 (placeId 0) with the third-party robloxstudio-mcp connected:

The fourth row is why selection cannot be clever. There is no waveform to analyse, no spectrogram to look at, nothing to judge by ear from here. What exists is a listing and a measurement, and the skill is honest about using only those.

The rule that shapes everything

A listing is a claim. Sound.TimeLength is a measurement. Never gate on a claim.

find_sfx.py deliberately stores the listing's duration under the name listed_duration so that nothing downstream can mistake it for data, and pick.py records listing_was_wrong on any pick where the two disagreed by more than 25%. On the first real run that fired on 2 of 8.

Two Studio calls, never one

execute_luau drops all output if the script yields at all, and a Sound does not resolve its length in the frame you set SoundId. So the probe is split:

Copy code
call 1   probe.lua    creates every candidate Sound, returns immediately
         (a beat)
call 2   report.lua   returns role,id,loaded,length as CSV

If anything is still loaded=false in call 2, run report.lua once more before believing it. Only a second failure is a real one, and a real one means an unresolvable id, a deleted asset, or audio this place is not permitted to use.

The three gates

A candidate is only eligible once Studio says all three, in order:

1. IsLoaded == true — the id resolves and the place may use it
2. TimeLength > 0 — a zero-length sound is broken, not short
3. min_len <= TimeLength <= max_len — the role's band

Search score breaks ties, and only ties. A top-ranked listing that fails a gate always loses to a bottom-ranked one that passes. On the first run this rejected a 6.55s "wooden gate open" from a role whose band ended at 5s.

An unfilled role is a failure, not a warning. emit_bank.py refuses to emit a bank with holes; overriding needs an explicit --allow-holes, which means deliberately shipping a silent event.

Writing a spec

One entry per thing the game needs a sound for, never per sound you imagine.

jsonCopy code
{
 "name": "farm",
 "roles": [
  { "name": "sheep_bleat", "query": "sheep bleat", "kind": "SoundEffect",
    "min_len": 0.3, "max_len": 4.0, "volume": 0.5 },
  { "name": "ambience",    "query": "countryside ambience", "kind": "Music",
    "min_len": 20, "max_len": 300, "volume": 0.2, "looped": true }
 ]
}

• kind is SoundEffect or Music. It is a real field on the asset and a mismatch is scored down hard, because a Music track answering a SoundEffect role is the classic way a nine-minute remix ends up as a footstep.
• The band is the design decision. A footstep over 1.5s is not a footstep, and writing that down is what lets the machine reject without hearing anything.
• volume and looped travel through to the emitted bank.

An example spec is in examples/spec.example.json; copy it and edit the roles.

Run it

bashCopy code
cd ~/.claude/skills/roblox-sfx
python scripts/find_sfx.py spec examples/spec.example.json --out work/candidates.json --per-role 5
python scripts/emit_probe.py work/candidates.json --out work
# execute_luau work/probe.lua      -> "created N sounds across M roles"
# execute_luau work/report.lua     -> save the return value verbatim to work/report.txt
python scripts/pick.py work/candidates.json work/report.txt --out work/manifest.json
python scripts/emit_bank.py work/manifest.json --out work
# execute_luau work/install.lua    -> SoundService.SFX, audible in this place
# execute_luau work/tour.lua       -> plays the whole bank in order, so a human can judge it

A one-off lookup, no spec needed:

bashCopy code
python scripts/find_sfx.py search --query "wolf howl" --kind SoundEffect --limit 20

What ships

Two rules the emitted code keeps:

• Nothing outside the bank ever names an asset id. Gameplay code says Sfx.play("wolf_howl", head). Retuning volume or swapping an asset is a data edit, never a code edit.
• Sound can never break gameplay. Every entry point in Sfx.luau is pcall-isolated and returns nil on failure. A missing sound must not stop a shot from firing or a sheep from moving.

tour.lua schedules with task.delay rather than looping with task.wait, for the same reason the probe is split in two: a yielding script loses its output.

Traps

The listing duration is wrong. Covered above. It is the single most likely thing to make this pipeline ship a bad sound if someone "optimises" the Studio round trip away.

search_assets on the MCP is a dead end. It returns HTTP 403 Scope not authorized, and no renaming of parameters fixes it. That is why this skill goes straight to the toolbox endpoints, which need no auth at all.

You cannot hear anything, and neither can the model. assetdelivery is 401 without a cookie, so there is no waveform, no spectrogram, no loudness figure. Everything this skill claims about a sound comes from the listing or from TimeLength. When taste matters, run tour.lua and let a human judge. For an unattended run, the gates plus the score are the honest ceiling, and that is a finding worth saying out loud rather than papering over.

Free packs are not the same as free audio. A free Creator Store pack is not loadable until a human takes it into the account (learned the hard way on the VFX harvester). Individual audio assets used by id have no such step, which is exactly why this skill works by id and never asks anyone to take anything.

Volume 0 during probing is deliberate. 40 candidate sounds resolving at once in an open Studio would be unpleasant. probe.lua sets Volume = 0; audition.lua, install.lua and tour.lua set the real volume.

Roblox's own stock library scores a deliberate bonus. Creator id 1 assets (the 91207…-shaped ids) are consistent, always permitted and never deleted. If a run must be reproducible months later, prefer them.

Why library selection and not generated audio

Selecting from the library needs no quota, no ID verification, no moderation wait and no upload, and an agent searching and selecting from the Creator Store needs no human intervention.

Generated audio would burn a create-only upload quota to produce something no one in the loop can hear. That trade is not worth making, and this skill is the argument.