agentleFS
eriknomitcheriknomitch/workmelt/.claude/skills/elevenlabs-sfx/SKILL.md

elevenlabs-sfx

Skill4 starsChanged 4 months agoMITView on GitHub

Generate a sound effect with ElevenLabs and ship it into the game through the existing SFX pipeline — masters in assets-src/eleven/, a curation entry in tools/sfx-sources.mjs, encoded Opus in public/sfx/, credits, and the wiring (if any) that makes the game actually play it. Use this skill whenever the user asks to create, generate or AI-generate a sound effect, replace or improve one of the game's sounds with a generated one, mentions ElevenLabs / /sound-effects / sound-generation for this game, or asks how a new sound gets into the audio system with the intent of adding one.

  • Reads credentials
  • Installs packages
  • Sends data out

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 1 is distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    If the user's request does not obviously map to a group/key ("make a cool explosion sound" — grenade impact?

    line 66
  2. Very specific:

    Use the windowed presets (SHOT, STEP, …) when the generation came back with dead air or a long tail.

    line 124
  3. Very specific:

    Add a CREDITS entry in the same file — rule 4 of AGENTS.md: attribution ships, and encoding an asset does not change its licence.

    line 131
  4. Very specific:

    Always filtered to the group(s) you touched.

    line 154
  5. Very specific:

    libopus is not byte-deterministic: Never --force; commit only the files for the group you changed and git checkout -- anything else that …

    line 159
  6. Very specific:

    Use ctx.rng, never Math.random().

    line 171
  7. Very specific:

    Credits ship: Update CREDITS in tools/sfx-sources.mjs.

    line 202

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

ElevenLabs sound effect -> shipped Workmelt sound

This skill takes a text description and produces a sound the game actually plays: generated masters in assets-src/eleven/, a curation entry in tools/sfx-sources.mjs, committed 48 kHz mono Opus in public/sfx/ with an updated manifest.json and CREDITS.md, plus whatever engine wiring the target needs — which is usually none.

The premise: samples overlay synthesis, they never replace it

The runtime synthesises every voice it needs (src/audio/weapons.js, foley.js, vox.js). src/audio/samples.js loads public/sfx/manifest.json in the background and AudioSystem._build() prefers a sample per (group, key), falling back to synthesis on any miss — a missing or undecodable file is silent degradation, never an error. So a generated sound is an upgrade to an existing hook, and the pipeline is the same one the CC0 packs and the authored reload foley already ride:

ElevenLabs API ─▶ assets-src/eleven/<key>_<i>.wav   (master, committed)
                       │  entry in tools/sfx-sources.mjs
                       ▼
              node tools/encode-sfx.mjs <group>
                       ▼
       public/sfx/<group>/<key>_<i>.opus + manifest.json + CREDITS.md
                       ▼
     src/audio/samples.js picks it up — zero engine code for existing kinds

Rule 3 of AGENTS.md stays satisfied: ElevenLabs is called offline, at authoring time. Nothing in the bundle fetches anything; the committed Opus is what ships.

Step 0 — route by target, and ask if it is ambiguous

Decide where in the game the sound plays before generating anything. Two cases, very different in cost:

Case A — the kind already exists (the goal; steer here). The target is an existing (group, key) that _build() dispatches on: a shot weapon profile, a step surface, an impact surface, a ui cue, a vox announcer line, a reload phase. Enumerate what exists from public/sfx/manifest.json and the GROUP table in src/audio/samples.js. Everything is curation-table-only.

Case B — a genuinely new sound. Needs a new kind: an emitter somewhere (ctx event → audio.play(kind, position, opts)), a GROUP entry in samples.js (jitter/send), a synthesized fallback voice (or an explicit, commented decision that this kind is sample-only and silent without assets), and a bus/priority choice. Read ARCHITECTURE.md and the header of src/audio/index.js first. This is real engine work — confirm scope with the user before doing it.

If the user's request does not obviously map to a group/key ("make a cool explosion sound" — grenade impact? new ambience?), ask rather than guess.

Step 1 — API key

The generation call needs ELEVENLABS_API_KEY. If it is unset or rejected (401), invoke the setup-api-key skill rather than improvising. The sound-effects skill has the full API reference; the one call this pipeline needs is below.

Step 2 — generate the masters

Generate into assets-src/eleven/ — these are our authored takes, committed like assets-src/vox|reload|shot (see step 3 for the gitignore exception). Request a high-quality format; everything is re-encoded to Opus anyway, so mp3_44100_192 is fine and pcm_44100 (headerless — wrap or just use mp3) is overkill:

curl -sf -X POST "https://api.elevenlabs.io/v1/sound-generation?output_format=mp3_44100_192" \
  -H "xi-api-key: $ELEVENLABS_API_KEY" -H "Content-Type: application/json" \
  -d '{"text": "<description>", "duration_seconds": 1.5, "prompt_influence": 0.4}' \
  --output assets-src/eleven/<key>_0.mp3
  • Variants matter. shot jitters playback rate by only 0.02 and step / impact pools exist to defeat repetition — generate 4 variants for those groups (re-run the call with reworded prompts; identical prompts converge). ui and vox take one.
  • Prompting: be specific and physical ("dry supersonic rifle crack, close mic, no reverb tail" beats "gunshot"). Ask for no reverb / dry — the mixer runs its own convolution reverb, and a baked-in room double-verbs (encoder header, item 3). For duration_seconds, aim near the window preset the group uses (SHOT 1.3 s, STEP 0.5 s, IMPACT 0.7 s, UI 1.0 s).
  • Listen before wiring (afplay on macOS). A generated take can be flatly wrong — burning an encode cycle on it wastes more than a regen costs.

Step 3 — commit the masters

assets-src/* is gitignored with per-directory re-includes. Add one for generated takes if it is not there yet, next to the existing exceptions:

!assets-src/eleven/

Then git add assets-src/eleven/. Committed masters are what make node tools/encode-sfx.mjs <group> work on a fresh clone without the 338 MB fetch-sfx.sh download — the encoder only needs assets-src/ to exist and warns MISS harmlessly on absent packs.

Step 4 — curation entry + credits

In tools/sfx-sources.mjs, add an entry under the right group with the right window preset. Read the presets at the top of the file; the choice that is not obvious: generated clips are usually already tight — there is no long take to hunt a transient inside — so { whole: true, fade: true, app: 'audio', bitrate: '64k' } (the RELOAD shape) is often righter than a peak window. Use the windowed presets (SHOT, STEP, …) when the generation came back with dead air or a long tail.

{ group: 'ui', key: 'mycue', src: ['eleven/mycue_0.mp3'], ...UI },

Add a CREDITS entry in the same file — rule 4 of AGENTS.md: attribution ships, and encoding an asset does not change its licence. Record it honestly:

{
  pack: 'ElevenLabs generated SFX',
  authors: 'Generated with ElevenLabs sound-generation (prompts in assets-src/eleven/)',
  license: 'ElevenLabs commercial licence (paid plan) — verify the account tier',
  url: 'https://elevenlabs.io/terms-of-use',
  used: '<which in-game sounds>',
},

ElevenLabs' free tier requires attribution and its paid tiers grant commercial use — surface the terms to the user rather than deciding licensing on their behalf.

Step 5 — encode, filtered, never --force

node tools/encode-sfx.mjs <group>

Always filtered to the group(s) you touched. A filtered run merges into the existing manifest.json instead of dropping groups it skipped, and MISS warnings for unfetched CC0 packs are expected, not failures. Two traps from the encoder's own header, both real:

  • libopus is not byte-deterministic. Never --force; commit only the files for the group you changed and git checkout -- anything else that shows modified, or the diff claims edits nobody made.
  • Needs ffmpeg on PATH (brew install ffmpeg).

Step 6 — wire (Case B only)

For an existing kind, there is nothing to do — the manifest entry is the wiring. For a new kind: emitter event per ARCHITECTURE.md, GROUP entry in src/audio/samples.js with jitter/send matched to the nearest existing group, bus choice in src/audio/index.js, and a synthesized fallback if the sound is gameplay-relevant (the game must never be silent because an asset is late). Use ctx.rng, never Math.random().

Step 7 — verify, including with your ears

node src/audio/attenuation.selftest.mjs   # routing: head-locked vs spatialised
node src/audio/reload.selftest.mjs        # if you touched reload foley
npm run build

Then listen in context — a clip nobody heard is the audio version of a .png nobody read. node src/audio/probe.mjs --port=<dev port> offline-renders every voice against the running dev server; or play the game to the moment the sound fires. Check: does it cut through the mix, does the double-verb trap show (tail too wet), does the variant pool loop audibly under repetition?

Step 8 — report

Say what was generated (prompts included), which group/key it landed on, what window preset and why, which checks ran, what you actually listened to, and the licence status. If you could not listen (no audio device), say so explicitly rather than implying a pass.

Hard rules

  • No runtime fetch to ElevenLabs, ever. Generation is authoring-time only.
  • Never --force the encoder; never commit re-encodes of groups you did not change.
  • Masters are committed in assets-src/eleven/; prompts are recorded (a prompts.md beside them, or in the curation entry's comment) so a take can be regenerated or iterated.
  • Credits ship. Update CREDITS in tools/sfx-sources.mjs; the encoder writes public/sfx/CREDITS.md from it.
  • Dry generations. The mixer owns the room; ask the model for no reverb.
  • assets-src/* is gitignored with narrow re-includes — a new master directory silently stays untracked until its !assets-src/eleven/ exception exists. Verify with git check-ignore -v <path> before assuming a file will be committed.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.