Addy Osmani's Agent Skills post on HN received 162 upvotes, and in the same week, seven of GitHub Trending's top ten were Agent-related—we judge that the competitive focus of Agents is shifting from "whose model is stronger" to "whose process constraints are better."
What This Is
SKILL.md is essentially a Markdown-formatted context injection (writing preset instructions into a file that AI automatically reads and follows when executing tasks). When you tell Claude Code "help me fix a bug," without SKILL.md, the AI modifies code directly and may introduce new issues; with SKILL.md, the AI first writes tests, runs tests to confirm reproduction, modifies code, runs tests again, and only commits after passing.
Key insight: model capability is exactly the same; the difference lies only in process constraints. The author's testing showed that when SKILL.md was over 200 lines, Claude Code's probability of skipping critical steps was about 30%; when shortened to 30 lines with Anti-patterns (i.e., explicitly forbidden wrong practices) annotated, the skip rate approached zero. A Skill is not documentation—it's instructions; the shorter, the more precise.
Multiple Skills can also be composed: when tdd.md, code-review.md, and debug.md are loaded simultaneously, writing code automatically triggers review, encountering a bug automatically triggers systematic debugging, and building new features automatically follows test-driven development. This is the prototype of Agent orchestration (multiple AI roles collaborating by rules).
Industry View
On GitHub, the obra/superpowers repository has already accumulated 40k stars building a similar Skill template library, and browserbase/skills is also pushing for standardization. The community is forming a consensus: prompts alone are insufficient; structured process files are needed to constrain Agent behavior.
But it's worth being cautious—the author himself ran into pitfalls: multiple Skills can conflict. tdd.md demands "refactor immediately," code-review.md demands "review before acting," and when triggered simultaneously, the AI doesn't know what to do. A more fundamental criticism is that this mechanism is essentially still a variant of prompt engineering—each additional Skill adds maintenance cost, and when models upgrade, previously effective Skills may need to be completely rewritten. Agent reliability issues can only be mitigated, not fundamentally cured, by process files.
Impact on Regular People
For enterprise IT: the Skill mechanism transforms AI coding assistants from individual tools into team-standardizable assets—teams share the same set of Skill files, ensuring everyone uses AI consistently.
For individual careers: programmers' work focus will further shift from "writing code" to "writing process constraints"—defining how AI should work is more valuable than doing error correction for AI.
For the consumer market: short-term impact is limited, but the logic is universal—if coding Agents can be constrained by process files, customer service Agents and writing Agents can too; a cross-industry Skill marketplace may emerge in the future.