SkillPedia Skills Encyclopedia
中文 EN

技能 2026-09-28 03:23 UTC

threejs-game-skillsv5.5

版本【5.5】由提交人 majidmanzarpour 提交,主要更新类型为:新增功能与能力扩展。提交说明:Tune skills for Claude Opus 5.5, cut QA token cost, and add music generation。带来了新的能力与可配置项,扩展了使用场景。

majidmanzarpour 8286774 majidmanzarpour/threejs-game-skills

Commit提交信息原文

Tune skills for Claude Opus 5.5, cut QA token cost, and add music generation

Audit the nine skills against current Claude behavior (Opus 5.5) and the
Opus 5.5 prompting guide, make the QA/screenshot pipeline faster and cheaper
in agent tokens, and fix two QA failures that cost agents failed runs.

Opus 5.5 prompt alignment:
- Director asks for a short update at each phase boundary instead of
  "only when" something important happens (current models under-narrate
  on long builds).
- Director names the early stops to avoid until the requested bar is met:
  ending a turn by announcing the next step, offering to continue, or
  listing decisions that don't block the work.
- UI patterns name the default web styles to avoid unless the art
  direction calls for them (cream panels, italic accent words, 01/02/03
  labels, monospace text labels, pill buttons), per the guide's advice
  that named patterns work and vague "avoid generic" lines don't.
- Normalize all-caps emphasis in the Tripo notes; every rule and its
  measured reason is kept.
- Replace the debug skill's 8-step script with the goal and the steps
  whose order matters; point to the playbook for triage order.

Director description:
- Replace the trigger-phrase list with request categories. In an 80-run
  headless Claude Code trial (Opus 5.5, 20 prompts x 2 runs per variant)
  the new wording loaded the director for 20/20 broad game requests
  versus 18/20, with no change to specialist routing (12/12) or
  unrelated prompts (8/8).

Contradictions and stale facts fixed:
- Audio reference no longer sets a minimum asset quota, matching the
  audio skill and the no-universal-quota rule.
- AGENTS.md requires mobile input only when mobile is a target,
  matching the UI and QA skills.
- Scaffold `--force` is described as it behaves: it copies into a
  non-empty directory and leaves other files in place.
- The "skills do not enable async APIs" note now lives only in
  asset-recovery.md.

QA and screenshot pipeline:
- `verify:visual` reuses a running dev server (reuseExistingServer
  outside CI) instead of exiting because the port is in use.
- The mobile Playwright project runs iPhone 13 emulation on GPU-backed
  Chromium (`mobile-chrome`) instead of WebKit, which the documented
  install never provided.
- Inspector `--manifest` captures every declared viewport/state in one
  browser and writes each report to its declared path (3 captures: one
  command, 3.6 s).
- Inspector prints one summary line per capture instead of the full
  JSON (3 captures: ~6.7 KB -> 594 B of agent-visible output); `--json`
  keeps the old behavior. Browser errors are counted in full but stored
  deduplicated, capped and truncated.
- Screenshots are saved at CSS scale (mobile 390x664 instead of
  1170x1992, about a fifth of the image tokens); desktop pixel metrics
  are identical and mobile within 2%.
- One screenshot per capture with the canvas cropped from it, `load`
  instead of `networkidle`, and no default settle wait for frozen named
  states.
- Install docs use `npx playwright install --no-shell chromium`.
- check_evidence.py's legacy discovery no longer walks node_modules,
  .git, dist or test output.

Token efficiency across the skills:
- The test-hook contract lives only in evidence-manifest.md; the QA
  skill, visual-test harness and bot playtest docs point to it.
- Remove repeated text: asset sourcing in the director and graphics
  builder (restated asset-recovery.md), debug profiling steps, Tripo
  checkpoint usage, image-generator lists in the 3D references, GPU
  benchmark history, and legacy checker details. About 1,150 tokens of
  duplication removed; net -400 tokens after the new content below.

Music generation:
- Audio helper adds a `music` command on ElevenLabs POST /v1/music:
  `--prompt` with `--duration` (3-600 s) and `--instrumental`, or
  `--plan` with a composition_plan JSON and `--seed`; invalid
  combinations are rejected before any paid request.
- Audio skill, reference, director routing and README cover music:
  a music row in the audio matrix, loop and stinger lengths, runtime
  looping and crossfades (the API has no loop flag), intensity layers,
  and mix balance.

Tests: 74 Python and 48 inspector tests pass (5 new for manifest mode,
settle waits, error caps and the summary line). Verified end to end on a
fresh scaffold: batch capture, manifest evidence check, verify:visual
with the dev server running, and the bot playtest.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>