技能 2026-04-28 08:43 UTC
music-composition
由提交人 SJY051 提交,主要更新类型为:常规更新与维护。提交说明:Polish benchmark caveats for external readers。对项目做了常规维护与小幅调整。
Commit提交信息原文
Polish benchmark caveats for external readers Four small edits to make the benchmark docs cleaner for a wider audience: 1. "Sei" agent reference: replaced with "the AI agent the author uses daily (Claude Code, Opus 4.7, Max tier)" / "스킬 작가가 평소 사용하는 AI 에이전트(Claude Code Opus 4.7 Max)" — outside readers don't know who Sei is. Author identity preserved in commit history and project context, but the public-facing methodology disclosure now reads without requiring background. 2. Single-model caveat: removed the specific Gemini reference (the region-restricted skills feature situation isn't documented well enough for the claim to be useful). Replaced with the cleaner general framing: "effect on other harnesses / services / models is theoretically expected (the skill follows the open Agent Skills standard) but unverified — cross-platform claim rests on format spec, not on tested behavior." 3. Cross-tier future-work item: corrected model versions. Latest Claude tiers as of run date are Opus 4.7, Sonnet 4.6, Haiku 4.5 (not Sonnet 4.5 as previously stated). 4. "No instrumentation" caveat: tool-call traces actually exist in Claude Desktop's thinking output, just impractical to extract across 8 prompts of dense Opus 4.7 thinking. Renamed "No instrumentation" → "Tool-call trace not extracted", body updated to reflect that instrumentation is observable in principle but wasn't extracted for this write-up. All four edits applied to both v1.0-claude.md and v1.0-claude.ko.md; edit (1) also applied to v1.0-eval.md and v1.0-eval.ko.md where Sei was mentioned in the methodology. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>