Date: 2026-08-28 13:15 Summary: Added a project Skill that cuts single-tool Agent token use via session splits and read allowlists without skipping coverage gates or i18n quality.
[question] 单独一个skill做token优化约束，但绝对不要降低生成工具的质量
[try to solve]
Created .cursor/skills/tool-token-efficiency/SKILL.md as a companion to tool-coverage-pass. It optimizes how agents work (A/B/C/D session boundaries at brief / master i18n / locales / ship; read only work-tasks/{slug}/ and slug shards; forbid merged src/site/i18n/{lang}.ts and full tool-catalog.json; reference pages via ≤150-line slices; filter lint output to the active slug; prefer npm run verify:tool -- --slug=).
Quality is non-negotiable and listed explicitly: 0b, 0i, phase 2/4, ten-locale search-oriented rewrites, ≥3 rounds, IG/FAQ/how_items/loadSample, and full ship gates. Conflict order: SEO/localization rules and coverage pass beat this Skill.
Registered in AGENTS.md, one-line pointers in tool-creation.mdc and tool-coverage-pass/SKILL.md, and .github/copilot-instructions.md.
Full analysis (causes, factor weights, tool-count conclusion): dev-logs/2026-08/2026-08-28-13-36-single-tool-agent-token-analysis.md.
[actions]
- Added
.cursor/skills/tool-token-efficiency/SKILL.md - Updated
AGENTS.md,.cursor/rules/tool-creation.mdc,.cursor/skills/tool-coverage-pass/SKILL.md,.github/copilot-instructions.md