agentleFS
eriknomitcheriknomitch/workmelt/.claude/skills/glb-weapon/SKILL.md

glb-weapon

Skill4 starsChanged 4 months agoMITView on GitHub

Turn a premade 3D model (.glb / .gltf, typically a downloaded gun) into a playable Workmelt weapon. Use this skill whenever the user drops a .glb or .gltf into the repo and wants it in the game, asks to add a gun, rifle, pistol, SMG or sniper from a downloaded or purchased model, asks how to load or import a 3D model as an asset, mentions Sketchfab / CGTrader / assets-src/models, or wants to swap or replace one of the four existing weapons with a model-derived one. Also use it for questions about how premade meshes get into this engine if the intent is to ship one.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 3 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Very specific:

    Instead, tools/glb-bake.mjs converts the model offline into a plain ES module of quantised typed arrays that src/weapons/models/ imports …

    line 39
  2. Distinctive:

    If the table alone is not enough, do not bake blind and squint at a render — project the geometry.

    line 105
  3. Very specific:

    Do not leave everything on one key.

    line 143
  4. Very specific:

    Commit the .data.js and the .materials.json.

    line 155
  5. Very specific:

    Create src/weapons/models/<id>.js returning the model contract (the full shape is in references/weapon-contract.md).

    line 161
  6. Distinctive:

    Is the optic slide-mounted?: If the optic sits directly on the slide's top face it recoils with it and belongs in the slide assembly.

    line 174
  7. Distinctive:

    If you are replacing a slot, keep the terminal ballistics of the weapon you replaced unless the user asks otherwise.

    line 204
  8. Very specific:

    Run the whole free sweep, not just the weapon suites — a new weapon id reaches into audio, and src/audio/reload.selftest.mjs is what …

    line 211
  9. Very specific:

    Do not hard-code a hex colour or a material anywhere.

    line 278
  10. Very specific:

    Use the seeded rng the subsystem hands you, never Math.random() — captures must reproduce.

    line 280

≈ “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

Premade model -> Workmelt weapon

This skill takes a .glb and produces a weapon the player can select, fire, reload and kill with: a committed geometry module, a builder in src/weapons/models/, a balanced entry in defs.js, and a green set of self-tests.

The premise: bake, because a weapon is static geometry

For a weapon, nothing is loaded from disk at runtime. That is a decision about weapons, not a blanket ban — see "Importing a 3D model" in AGENTS.md. A rigged, animated character is the other case and takes the GLTFLoader + MeshoptDecoder path, because the bake discards skins and clips by design.

A weapon is static single-material geometry whose animation comes from clips.js driving named attachment nodes, so it loses nothing to the bake and gains everything below. If you are here for a gun, bake it.

AGENTS.md forbids runtime dependencies and CDN fetches — "every asset the game needs ships in the bundle, so it runs fully offline" — and assets-src/* is gitignored because the master is never the shipped artefact. Baking satisfies that with no loader in the bundle at all, which is strictly cheaper than the alternative for geometry that carries nothing a loader would preserve.

Instead, tools/glb-bake.mjs converts the model offline into a plain ES module of quantised typed arrays that src/weapons/models/ imports like any other source file. This is the SFX pipeline applied to geometry: the untracked master lives in assets-src/, a tool converts it, and the output is what gets committed.

What that buys, concretely: no loader in the bundle, no fetch, no parse cost on a boot TEXTURE-PERF.md already measures at 3.27 s of blocking work, determinism for tools/baseline.mjs, and a diff that shows part names, materials and triangle counts as plain text.

The model is geometry and nothing else. Its materials, UVs, tangents, skins and animations are all discarded — the engine's look comes from the procedural material library plus the curvature-mask bake, and its animation comes from clips.js driving named attachment nodes. Supplying those nodes is the real work of this skill, not the mesh.

Reference files — read before writing code

  • references/weapon-contract.md — the model return shape, the complete node table (required vs per-mechanism), the material keys, the defs.js balance contract, the LOADOUT decision, and every verification command. Read it in full before authoring the builder.

src/weapons/models/g31.js is the worked example — the first weapon built this way, end to end: the bake command in its header, the assembly split, nodes derived from bakedBounds, a generated magazine standing in for one the source lacked, and honest UNVERIFIED notes where a static capture could not prove a number. Read it alongside src/weapons/models/g31.materials.json. Start there.

Also read src/weapons/models/pistol.js end to end — it is the shortest fully procedural example of the contract, and it is what g31.js was derived from. src/weapons/models/baked.js documents the runtime API (addBaked, requireBaked, bakedBounds). ARCHITECTURE.md covers the engine contract if you need to touch anything outside src/weapons/.

Workflow

1. Inspect the model before touching anything

node tools/glb-bake.mjs assets-src/models/<file>.glb --id=<id>

The tool reads the binary container only. A .gltf + .bin pair must be converted first — gltf-transform copy in.gltf out.glb, or re-export as glTF Binary from Blender. Downloads arrive in both shapes, so check before assuming.

With no --out this writes nothing and prints a part table plus a material-map scaffold. Read it carefully — everything in step 2 comes from it:

  • Triangle count. Budget: the four shipped weapons are 8k–24k triangles. Past ~30k, say so and propose decimating in the source tool first.
  • Extent. The engine works in metres. A real handgun is ~0.20–0.26 m long, a carbine ~0.75–0.90 m. If the extent is 100× or 0.01× that, the export is in centimetres and needs --scale.
  • Which axis is long. The engine points weapons down −Z. Sketchfab exports routinely point down +X or +Y.
  • Groups before part names. Exports name every leaf Object_12, but the PARENT nodes are usually named by whoever built the model — RMR_0, G31_1, RMR CUT_2 — and that is exactly the split a weapon wants. The group column carries it, and selectors match either name, so prefer include: /RMR/ over enumerating leaves that a re-bake could renumber.
  • Centres and extents disambiguate the rest: a 0.223 m part at slide height is the slide, a 0.002 m-thick blade inside the trigger guard is the trigger.

If the table alone is not enough, do not bake blind and squint at a render — project the geometry. A ~60-line throwaway in scratch/ that scanline-fills each part's triangles into an ASCII side elevation (across = Z, up = Y) makes the slide, grip, trigger guard and optic instantly obvious, costs one round trip, and needs no browser. That is how the G31 was identified.

2. Decide placement, then the material map

Placement. Compose --rot, --scale and --origin until the model sits in engine space: muzzle down −Z, bore axis on +Y up, origin at roughly the web of the firing hand (pistol.js puts the bore 36 mm above origin). --origin is applied after --rot/--scale, so read the printed bounds, then subtract. Iterate on the inspect command — it is instant and writes nothing.

Aim the placement at the weapon you are replacing. Landing the bore height and the slide/receiver rear on the same numbers as the existing weapon is worth real effort, because it makes that weapon's solved hand targets a legitimate starting point instead of a guess. The G31 bake picks --origin=0,0.303,0.0316 for exactly this: bore at y 0.036 and slide rear at z 0.052, both the P-19's. Everything downstream of §4.3 gets cheaper.

Materials. Save the printed scaffold next to the baked output as src/weapons/models/<id>.materials.json and map every part to an engine key. Committing it keeps the re-bake command in the generated file's header reproducible. Keys resolve most specific first — "<group>/<node>", "<node>", "<group>", then the glTF material name — because a modeller reuses a material for whatever was the same colour in the viewport. On the G31 one Material.001 covers both the red-dot housing and the polymer grip, so material-name keys alone cannot express it. Keys starting with _ are notes.

Three things that are not guesses (all detailed in the contract doc §5):

  • The optic lens is glass. Any other key gives you a sight that looks perfect from outside and is opaque the instant you aim through it. Nothing warns you; only ?view=ads shows it.
  • Prefer steel_soot over steel_black for an imported slide. steel_black has the library's highest wear amplitude (0.24) and the wear mask is per-vertex, so on faceted geometry it bleaches whole panels white.
  • Do not leave everything on one key. The polymer/alloy/rubber split across frame, receiver and grip is most of what sells it as a real object.

3. Bake

node tools/glb-bake.mjs assets-src/models/<file>.glb \
  --id=<id> --out=src/weapons/models/<id>.data.js \
  --map=src/weapons/models/<id>.materials.json \
  --rot=0,90,0 --origin=0,0.303,0

Commit the .data.js and the .materials.json. Never edit the .data.js by hand — it is generated, and its header carries the exact command to reproduce it.

4. Author the builder

Create src/weapons/models/<id>.js returning the model contract (the full shape is in references/weapon-contract.md). The order that works:

  1. Split the assemblies. Body first, then one Assembly per moving part using requireBaked with a selector over the part node names. A slide, magazine and trigger that stay in the body can never animate. Use requireBaked, not addBaked — a selector that matches nothing is this pipeline's signature silent failure. Author the BODY as an exclusion, so a part added by a future re-bake lands on the frame — visible and wrong — rather than silently vanishing.

    Two things to check in the source before splitting, both common:

    • Is the optic slide-mounted? If the optic sits directly on the slide's top face it recoils with it and belongs in the slide assembly. Left on the frame it hangs in the air while the slide cycles underneath.
    • Is there a magazine at all? Often not — the grip and magwell are one solid part with nothing below the grip floor. Check whether any part's min-Y reaches past the grip. If none does, generate one with buildMagazine from parts.js, exactly as pistol.js does.
  2. Derive nodes from geometry, not guesses. bakedBounds(PARTS, /slide/) gives you the real numbers for muzzle, ejection port, sight and magazine seat. Write the derivation down in a comment.
  3. Hand targets last, and expect to iterate. gripR/gripL are wrist targets, not palms, derived as knuckle − 0.098 × finger. They are the one thing that cannot be got right from numbers alone — see the contract doc and the long derivation comments in rifle.js.

5. Wire it into the game

A mesh that renders is not a weapon. Read the LOADOUT section of references/weapon-contract.md and decide, with the user, whether this replaces one of the four slots or extends the loadout — the two paths differ in what else has to change. Then add the defs.js entry against the balance contract (a shots-to-kill number, not a feel number) and register the builder.

The weapon id is a key into other subsystems, and all of them fail quietly. Contract doc §7 has the table; the ones that bite are FOLEY_ALIASES in src/audio/samples.js (reload foley, else all four phases drop to synthesis) and the resolveProfile regex in src/audio/weapons.js (else the gun fires with the wrong shot profile). Both are keyed on the hardware, so a new gun that is mechanically an existing one should alias to it rather than get its own.

If you are replacing a slot, keep the terminal ballistics of the weapon you replaced unless the user asks otherwise. The slot's damage and rpm ARE the balance contract; a new model is a reason to change how a gun handles, not how hard it hits. Spend the character on recoil, bloom and handling instead.

6. Verify

Run the whole free sweep, not just the weapon suites — a new weapon id reaches into audio, and src/audio/reload.selftest.mjs is what catches it:

for f in $(find src server \( -name 'selftest.*' -o -name '*.selftest.*' \) \
    | grep -v '^server/' | grep -v 'src/audio/selftest.js' | sort); do
  node "$f" >/dev/null 2>&1 || echo "FAIL $f"
done
npm run build

Then look at it, which for a weapon is not optional — every remaining failure mode in this pipeline is visual. Use the visual-check skill, and the weapons preview harness, which exists for exactly this:

/src/weapons/preview.html?w=<id>&view=hero    # also: side, muzzle, optic, grip, hands, ads, reload

tools/capture.mjs cannot shoot this page (it is --shot= registry-driven), so drive it with a small playwright script in scratch/ — one page per view, ~4.5 s settle for the material bake, screenshot to .shots/<id>-<view>.png. Listen for pageerror while you are there; a builder that throws still leaves you a plausible-looking sky.

?view=ads is mandatory, not optional. It is the only view that shows whether the optic is transparent, and an opaque lens looks perfect in every other view. Check in this order:

  1. side — silhouette, scale and material split against the existing guns
  2. ads — can you see through the sight? plus the reticle
  3. hero — how it reads in the round
  4. grip / hands — hand placement, expect to iterate

Know what a static view cannot prove. A seated magazine is entirely inside the grip, and ?view=reload samples the clip at t=0 before anything moves, so magSeat is not verifiable from these captures — say it is underived rather than implying it passed. Use &t=<seconds> to sample mid-clip if you need it.

7. Report

Say what the source model was, its triangle count, which parts moved into which assembly, how the hand targets were derived, which checks ran, and which views you actually looked at. If you skipped a render because the environment had no GPU, say that explicitly rather than implying a pass.

Hard rules (the ones that are easy to break silently)

  • Never add a runtime loader or fetch a model file. The bake is the only path. No GLTFLoader import in src/.
  • Never hand-edit a .data.js. Re-bake instead; the header has the command.
  • Never commit the source .glb outside assets-src/ — it stays untracked, like the SFX masters.
  • requireBaked, not addBaked, for every moving part. An empty assembly renders perfectly and silently stops the gun cycling.
  • Check the licence FIRST, before any other work — it can invalidate the whole task. assets-src/models/*/license.txt. Baking geometry into source does not dissolve a CC-BY obligation: the credit ships. Add it to public/models/CREDITS.md (the established location, mirroring public/sfx/CREDITS.md) and surface the terms to the user rather than deciding licensing on their behalf.
  • The optic lens is material key glass. Not alu_fine, not steel_bright. Anything else is opaque under ADS and silent everywhere else.
  • The weld keys on position and normal. If you ever touch it: a position-only weld smooths every hard edge on a low-poly gun and looks like a material bug. baked.selftest.mjs guards this.
  • Do not hard-code a hex colour or a material anywhere; surfaces come from the keys in src/weapons/materials.js.
  • Use the seeded rng the subsystem hands you, never Math.random() — captures must reproduce.

New files this skill adds to the repo

.claude/* is gitignored with narrow re-includes, so new skill files need git add -f. Existing tracked ones keep tracking. 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.