32 Skills installed, 17 ran in vain — a backend engineer recapped three months of hard lessons on Juejin, sending a wake-up call to the entire AI tool community: precise beats plentiful.

What This Is

Claude Code's Skill system is essentially a "plugin" marketplace for the AI assistant. Each Skill is a markdown file with two parts: the opening tells the AI "when to trigger," the body tells it "how to act." In theory, you can install dozens without conflict — but Claude Code's context budget (think of it as the AI's "working memory") reserves only 1% by default for all Skill descriptions. Install more than 30, and the descriptions at the back get auto-truncated. The AI won't even know the tool exists.One change worth watching in this mechanism: SKILL.md has become a cross-platform open standard. The same Skill file runs on Claude Code, Cursor, and Gemini CLI. Write once, deploy everywhere.

Industry View

The community compares the Skill ecosystem to "npm package management in the AI era" — by early 2026, several core repositories had collectively crossed 60K stars and 2,000+ usable Skills. The ecosystem is moving fast. But we see three red flags:First, quality review is missing. The most comprehensive repo (37K stars) comes from a third-party team, not Anthropic official — the official repo has only about 20 Skills. Beneath the prosperity, sand runs with the gold. This is exactly the script npm played out years ago.Second, "raising the budget" treats the symptom. The article mentions bumping the budget ratio from 1% to 2%, but the root issue is that 30 Skills are inherently redundant. Tweaking parameters won't hide a structural problem.Third, fragmentation risk. Superpowers, Vercel Labs, and various third-party teams are all building "complete suites." Incompatibility between them is a matter of time.

Impact on Regular People

For enterprise IT: rather than reporting "how many AI tools we installed," first audit which workflows are actually used. Reuse rate is a better metric than tool count.For individual careers: a universal lesson — in any tool stack, 10 high-frequency items cover 90% of scenarios. The rest is mostly a placebo sitting in your bookmarks.For the consumer market: the AI assistants and smart customer service bots on your phone face the same problem — the more features pile up, the harder users find the entry point. The next wave of AI product competition won't be "add features." It will be "do the subtraction."