levy-street/world-of-claudecraft / .claude/skills/asset-pipeline
asset-pipeline
VerifiedAsset pipeline: generate, review, integrate
Full reference: scripts/asset_pipeline/CLAUDE.md. This is the operational loop.
0. Prerequisites
TRIPO_API_KEYin the repo-root.env(gitignored).OPENAI_API_KEYoptional (better concept images, skin--promptrepaints). Never commit.envor keys.- Confirm credits before spending:
node scripts/asset_pipeline/pipeline.mjs balance. A P1 image-to-model run is ~40 to 50 credits; rig 25; retarget 10 per animation.
1. Pick the lane
| Ask | Lane | Command core |
|---|---|---|
| Held weapon (sword, axe, staff, ...) | weapon | weapon --name <key_with_family> --prompt "..." |
| World object / building / scenery | prop | prop --name <key> --height <units> --prompt "..." [--building] |
| Mob / NPC / animated character | creature | creature --name <key> --prompt "..." |
| Player-class skin (texture swap) | skin | skin --class <cls> --suffix <x> --tripo --prompt "..." |
| League-style themed CHARACTER skin | skinmodel | skinmodel --class <cls> --theme "pool party" --name <key> |
| Free rig of a raw mesh (KayKit skeleton) | rig-manual | rig-manual --raw <job>/raw.glb --name <key> |
skinmodel is the flagship: it renders the REAL base class model, has gpt-image-2 redesign
that exact character around the theme (same identity, chibi proportions), builds it with
Tripo's best model (v3.1 + smart_low_poly), retargets the FULL KayKit clip vocabulary
in-place (so clips: kaykit([...]) drives it unchanged), and injects calibrated
handslot.r/.l bones (pose transplanted from the knight reference) so it holds weapons
through the game's own attach path. Review preview/held_attack.png to see it swing a sword.
Weapon keys MUST contain a family token (sword, dagger, staff, hammer, axe, halberd, spear,
scythe, wand) or the tests/held_weapon_models.test.ts contract fails. For skins,
--tripo --prompt "..." is the RECOMMENDED mode (real re-texturing); --recolor is the
cheap deterministic fallback, and a bare --prompt repaint needs OPENAI_API_KEY
(mode details: scripts/asset_pipeline/CLAUDE.md).
2. Generate, then ALWAYS review the previews
Run the lane command WITHOUT --apply first. Then Read (the Read tool, they are images) the
preview PNGs under tmp/asset_pipeline/<job>/preview/: front.png, right.png, back.png,
left.png, hero.png, plus clip_<Name>.png per animation for creatures. Check:
- Weapon: blade/tip up (head up for axe/hammer/staff), looks like the request; the preview
dir carries
held_hero/right/attack.png(knight) ANDheld_<model>_{hero,right,attack}.pngfor ALL 7 class bodies with mid-attack frames: a weapon must hold correctly on every character. In the live viewer (library --serve) use the "held by" dropdown to watch any character (including generated skinmodel bodies) swing it. - Prop: upright, front facing the camera in
front.png. - Creature: every clip frame posed (a T-pose clip frame means a broken retarget); for non-biped rigs the walk clip is reused for Idle/Run/Attack/Death, judge if that reads OK.
3. Fix by resuming the job (paid stages never re-run)
- Weapon upside down: rerun the same command with
--flip --job <id>. - Prop facing the wrong way: rerun with
--rotate-y <deg> --job <id>. - Crash or timeout: rerun with
--job <id>(thejob.jsonledger skips finished steps). - Force specific steps to re-run after a parameter change:
--redo <step1,step2> --job <id>(e.g.--redo retarget,assemble); cleared paid steps re-pay, so prefer the targeted flags. - Wrong shape entirely: new run with a sharper
--promptor a--imageconcept (T-pose reference for creatures). This is a new paid generation; checkbalance. status [--job id]lists jobs;validate/inspect/inplace-checkre-check a GLB.- Lane and viewer detail beyond this loop (the
libraryviewer and--servelive 3D inspector,skin --triporeal re-texturing vs the--recolorfilter,skinsetwhole roster sets and their SKINS/SKIN_COUNTS lockstep) lives in the full reference,scripts/asset_pipeline/CLAUDE.md; read the matching section there before running those lanes. One rule worth repeating: skins are texture swaps on the SAME rig; a radically new body silhouette is the cosmetic-body path (SkinCatalogunion change, not automated by the skin lanes).
3b. QA gate (mandatory before apply)
node scripts/asset_pipeline/pipeline.mjs qa --job <id>
Lane-aware structural re-verification (rig + required clips, grip convention + HUD icon +
held-on-all-7-characters renders for weapons, handslots + KayKit vocabulary for skinmodels,
preview coverage) plus the REAL itemized cost: each recorded Tripo task is priced from the
API's own credits_consumed (1 credit = $0.01) and gpt-image-2 usage at token rates.
PASS/WARN/FAIL scorecard; qa.json in the job dir; exits 1 on FAIL (fix with --redo or a
regenerate before integrating). ALWAYS report the printed TOTAL cost per asset to the user.
4. Apply, guard tests, manifest
Rerun with --apply (still --job <id> so nothing regenerates). Then:
- weapon:
npx vitest run tests/held_weapon_models.test.ts - skin:
npx vitest run tests/skin_event.test.ts - creature/prop: place the printed snippet (step 5), then
npx tsc --noEmit - All lanes:
node scripts/build_media_manifest.mjs generate(auto innpm run build; dev needs no regen). Never hand-editsrc/render/assets/manifest.generated.ts. - CREDITS.md was auto-appended by
--apply;npm run asset:budgetis advisory.
5. Manual follow-ups per lane (the pipeline will not do these)
- weapon: place the printed ItemDef snippet in
src/sim/content/items.ts(or map existing item ids via--items); real vanilla-style stats are your judgment. - prop: add the printed
PROP_ASSET_DEFSentry (src/render/props.ts) and place it, eitherZonePropsDefinsrc/sim/content/zone*.tswith a collider matched to the visuals (src/sim/colliders.ts), or theGROUND_OBJECTSinteractable lane (no collision). - creature: add the printed VisualDef to
VISUALSand wire the mob template inMOB_KEYS(src/render/characters/manifest.ts); set the real world-unit height and tint. - New player-facing entities also need:
src/ui/world_entity_i18n.tsname entries and wiki regen (npm run wiki:content,npm run wiki:stillsfor models). Every new ITEM id also owes committed WebP art (tests/item_art_consistency.test.ts) in the same change, and a wordy English name owes its M16 non-Latin fills too (rootCLAUDE.mdi18n bullet); "English only at PR tier" does not cover those two. - Never extend
SkinCatalog(src/sim/types.ts): it is a closed sim/wire union. Class variants go through the skin lane; new bodies are mobs/NPCs. - Verify in game:
npm run dev, screenshot the asset in place.
6. Commit
Conventional Commits with a scope, e.g. feat(assets): add emberfang sword weapon variant.
Commit the GLB/texture/icon under public/, the registry edits, CREDITS.md, and your
snippet placements. Never commit tmp/asset_pipeline/ artifacts, .env, or keys.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |