Matt Pocock's Skills Repo — Engineering + Productivity Skills

Source: raw/gh-star-mattpocock-skills.md (GitHub star stub, 2026-05-20) + README extraction from github.com/mattpocock/skills.

Matt Pocock’s skills is a public mirror of his personal .claude/ directory — “Skills for Real Engineers.” Eighteen skills span engineering (diagnose, tdd, triage, improve-codebase-architecture, prototype) and productivity (caveman ultra-compressed mode, grill-me planning interviews, handoff session summaries). The pack is explicitly designed to address the four documented failure modes of AI coding agents: misalignment, verbosity, broken code, and architectural decay. MIT license, 96.4K GitHub stars at 2026-05-19 capture, primary language Shell.

Key Takeaways

• 18 skills total, split into engineering / productivity / misc.
• MIT license, maintained by Matt Pocock (well-known TypeScript educator).
• 96.4K stars at 2026-05-19 — exceptional for a personal-dotfiles-style repo, consistent with Matt Pocock’s existing TypeScript-community following.
• One-line install: npx skills@latest add mattpocock/skills. Then run /setup-matt-pocock-skills to configure issue tracking, labels, and documentation paths.
• Explicit failure-mode framing. The README’s load-bearing claim is that these skills target four specific LLM coding-agent failure modes: misalignment (agent does the wrong thing), verbosity (agent buries the answer), broken code (agent ships syntax errors), architectural decay (agent makes the codebase worse over time).
• caveman skill is the standout productivity move — ultra-compressed communication ~75% token reduction. Distinct from token-efficiency tricks like /compact because it changes Claude’s output style, not the session compaction layer.

The 18 Skills

Engineering (10)

• diagnose — Structured debugging for complex bugs. Multi-step diagnosis frame.
• grill-with-docs — Deep planning that updates domain-language docs as it goes.
• triage — Issue triage state machine.
• improve-codebase-architecture — Identify design improvements.
• tdd — Test-driven development loop.
• to-issues — Convert specs into GitHub issues.
• to-prd — Generate PRDs from conversations.
• zoom-out — Request broader code context.
• prototype — Build throwaway prototypes.
• setup-matt-pocock-skills — Initial configuration.

Productivity (4)

• caveman — Ultra-compressed communication (~75% token reduction).
• grill-me — Intensive planning interviews. Forces decisions out of the human via Socratic pressure.
• handoff — Compact conversation summaries for cross-session continuity.
• write-a-skill — Create new skills (the meta-skill).

Miscellaneous (4)

• Git guardrails
• Test migrations
• Exercise scaffolding
• Pre-commit setup

Implementation

Tool/Service: mattpocock/skills (personal .claude/ mirror) Setup:

npx skills@latest add mattpocock/skills
# then in Claude Code:
/setup-matt-pocock-skills

Cost: Free (MIT). Integration notes: Installs to ~/.claude/skills/ via the skills CLI. The /setup-matt-pocock-skills slash command configures issue-tracker integration, label vocabulary, and documentation paths — this is the one-time setup gate before the engineering-side skills are useful. Primary language Shell suggests skill internals lean on shell pipelines rather than Node/Python wrappers.

Differentiation vs Adjacent Frameworks

Tool	Scope	Notable skill	What’s distinctive
mattpocock/skills	18 personal-dotfiles skills	caveman (75% output compression)	Failure-mode-targeted; explicit alignment + verbosity + correctness + architecture frame
gstack	23+ role personas	/cso (security)	Role-persona model
GSD	Six-step loop	/gsd-execute-phase parallel waves	Phase-loop + artifact layer
Superpowers	Closed full-SDLC methodology	Comprehensive	Whole-lifecycle bundle
Six Best Business Skills	Per-business-task curation	Document/research focus	Business outcomes orientation
Marketing Skills Bundle	Marketing workflows	Campaign tooling	Marketing-specific

Closest peers are Six Best Claude Code Skills and Seven Skills That Run My Business — both are curated personal-skill collections shipped as references. mattpocock/skills is engineering-focused; the business collections are business-task-focused.

The caveman Skill (Detail)

The caveman skill is the most distinctive deliverable in the pack. It changes Claude’s output style to ultra-compressed bullet/keyword format, claimed ~75% token reduction. Distinct from:

• /compact (compresses session context, not output)
• Token-limit overrides (cap on output, not compression)
• “Be concise” system prompts (Claude tends to revert)

The caveman mechanism is a system-prompt + output-pattern enforcement. Useful when Claude’s verbose default is the bottleneck — e.g., diff explanations, large refactor plans, post-mortems. ^[inferred — the exact prompt and reduction measurement methodology aren’t in the README extraction]

Open Questions

• Star-count provenance. 96.4K stars is very high for a personal-dotfiles repo. Worth periodic check.
• caveman measurement methodology. ~75% token reduction is the headline number — how is it measured? Same task, two runs, before/after token count? Or estimated? Worth confirming before citing externally.
• Shell-primary language. Repo is Shell-primary. Worth checking how skill files are structured — pure shell? Mixed shell + markdown skill definitions?
• Failure-mode-targeted claim. “Designed to address AI coding agent failure modes: misalignment, verbosity, broken code, architectural decay” is the README framing. Worth verifying which skill maps to which failure mode (caveman ↔ verbosity is obvious; the others are less clear).
• Update cadence. Last commit / push date wasn’t surfaced in the WebFetch summary — typical follow-up before declaring a skill pack a stable dependency.

Related

• Claude AI topic
• Six Best Claude Code Skills for Business
• Seven Claude Skills That Run My Business
• gstack — Garry Tan’s 23-Tool Setup
• Superpowers — Agentic Skills Framework
• Skill Design Patterns
• Anthropic Engineers’ Four Skill Rules
• Agent Skills Overview
• 18 Claude Code Token-Optimization Techniques — caveman is a complementary lever to the techniques here.

Try It

1. Install in a sandbox profile before merging into your daily Claude Code setup — CLAUDE_HOME=~/sandbox-claude npx skills@latest add mattpocock/skills.
2. Run /setup-matt-pocock-skills in the sandbox and walk through the configuration. Decide whether the issue-tracker integration and label vocabulary match your project setup.
3. Test caveman on a verbose-default task (e.g., a refactor plan or code-review pass). Measure the actual token reduction on your workload — claimed 75% but worth verifying.
4. Use grill-me for spec-clarification before starting any 3+ hour project. The Socratic pressure to surface unstated assumptions is the highest-leverage move in the pack.
5. Steal individual skills rather than bulk install if the failure-mode framing doesn’t match yours — the skills are MIT-licensed, so cherry-picking caveman or handoff into your existing setup is fair game.