Managing 300+ Skills: Author Cuts Startup Context by 89% with a Skill-Slimming Workflow

卡尔的AI沃茨 · wechat · 2026-08-05

Author Karl shares his experience managing 300+ Skills in Claude Code and Codex. He found that each new session loads all Skill names and descriptions, consuming 9.9k tokens of startup context, vs 1.5-1.7k with custom Skills disabled—an overhead of 8.2-8.4k tokens. Based on July usage, this wasted an estimated 400-500M tokens of context space.

He categorized Skills by usage frequency: 22 high-frequency core, 64 project-specific, 214 low-frequency archive, and 21 RARECRITICAL. A review revealed 147 still globally discoverable. He built a local web tool, SkillScope Console, with search/filter/classification and introduced a "trigger state" intermediate: only a trigger word is exposed, and the full Skill is dynamically loaded when hit, saving 89% of startup context per Skill.

He packaged the governance process into a "Skill Slimming" Skill with five phases. As of Aug 4, 2026, out of 318 unique Skills, 147 are globally discoverable (29 core + 118 conservative), 14 project-level, 157 trigger-level, and 0 deleted. He emphasizes that Agent capabilities remain intact while model intelligence is freed, and reflects that "when skills pile up, judgment is more valuable than the skills themselves."

Original post →

More from coding & agent

coding & agent channel →