5 Claude Skills from Matt Pocock's Library

The Breakthrough

Sean Kochel demonstrates five skills from Matt Pocock's open-source Claude skill library that systematically improve AI-assisted coding for intermediate builders by addressing architecture, planning, efficiency, context, and session continuity.

What Actually Worked

• Improving Codebase Architecture dispatches an exploration agent through the codebase to identify architectural friction, then lists the top five deepening opportunities in priority order; each includes related files, the fundamental problem, a concrete solution, and benefits like better testability and AI navigability.
• Grill Me stress-tests a proposed change through 7-10 iterative questions that spiral down design branches, resolving specifics like problem shape, impacts on interacting functions, and contents of new modules before generating an implementable design.
• Caveman forces concise 'smart caveman' responses that drop fillers, articles, and pleasantries while preserving technical accuracy; auto-exits for security warnings, irreversible actions, multi-step sequences, or user requests, claiming up to 75% token reduction.
• Zoom Out analyzes domain vocabulary, involved modules, read/write files, and higher-level context to explain a code section's fit in the bigger picture, grounding explanations in project-specific language.
• Handoff distills the conversation into a markdown brief with problem framing, solutions, key decisions, and specifics; users provide instructions like 'pass to spec-driven tool for implementation' to bridge to fresh context windows.

Before / After

Caveman mode reduced a sample response from 768 tokens (Opus tokenizer) to 502 tokens, a roughly 30% cut; library claims up to 75% overall.

Context

Vibe coders—exit-level beginners to intermediates building systematically—risk convoluted codebases from inconsistent changes. Kochel demos the skills on his BMAD-built Twitter intelligence tool for trend identification, clustering, and ranking. The open-source library integrates with tools like OpenSpec and spec-driven development, surfacing high-level issues, ensuring human-resolved decisions, saving tokens, clarifying context, and porting plans without losing details.

Notable Quotes

• "It makes sure that it's actually using the language of your app to explain things to you."
• "Once you pick a direction it's going to go deep down the rabbit hole to resolve all of the other issues that crop up because of that decision."
• "Plan claim ranker recomputee threshold and divergence makes tuning log a lie."
• "This is kind of like an alternative to compacting because we're still going to get all of that information but we can then just use that document as the context for our next session."

Content References

No external books, papers, reports, podcasts, datasets, or events are specifically referenced beyond the core library.