Matt Pocock, the TypeScript authority with 60,000 subscribers, just open-sourced his private Claude Code skills workflow this week. His core thesis: when AI coding agents go off the rails, the model is to blame roughly 10% of the time—the other 90% is on us for failing to think through requirements. He ships a "force clarity before acting" skills stack: from /grill-me that interrogates your requirements, to /tdd that enforces test-first development, to /to-spec that turns conversation into specification documents.
What this is
Matt categorizes AI coding failures into four buckets: not doing what you wanted, being too verbose, not running, or writing spaghetti code. Each has a corresponding skill: the first two rely on /grill-me and /grill-with-docs to force you to think clearly—the latter also captures decisions as ADRs (Architecture Decision Records, i.e., "why we designed it this way" trail files); the latter two rely on /tdd plus /diagnosing-bugs, and /codebase-design.
The design splits into two layers. User-invoked acts as the "dispatcher"—you manually trigger /xxx to direct the flow; model-invoked acts as the "execution worker"—the agent calls these automatically to enforce specific disciplines. Installation is a single command: npx skills@latest add mattpocock/skills, then run /setup-matt-pocock-skills to align your issue tracker (bug tracking systems like GitHub Issues) and documentation location—this step can't be skipped, or subsequent skills won't know where to land.
Industry view
Supporters see Matt's version as a direct revolt against "vibe coding (coding by gut feeling without thinking it through)" and a course correction from heavyweight frameworks like GSD and BMAD—small, modifiable skills fit domestic teams better.
But there's pushback. The developer community points out these skills systems are essentially "using prompt engineering to pay off requirements engineering debt"—as projects grow complex, CONTEXT.md (shared project glossary) quickly turns into a dump; others criticize Matt's default GitHub+Linear ecosystem as ill-suited to the domestic market. This has spawned more aggressive domestic alternatives like spec-superflow—which front-loads mandatory specifications, at the cost of a steeper learning curve.
Impact on regular people
For enterprise IT: If teams start using skills, the "AI coding assistant" is upgrading from chat tool to workflow tool. IT leads need to start worrying about skill maintenance, version sync, and who maintains the shared glossary governance.
For individual careers: The key skill for using AI well is migrating from "can you write prompts" to "can you break down requirements." Even non-product managers will find the "force clarity before acting" mindset increasingly valuable.
For the consumer market: No immediate short-term impact, but once the open-source skill ecosystem matures, low-code platforms (tools that let non-programmers build systems) will faster make "clarify requirements first" a standard button.