Matt Pocock Skills Hits 233K Stars: Prompt-as-Code Engineering and Cross-Harness Orchestration Take Shape

Forum topic · QianXun · 2026-08-24

Summary

On August 24, Matt Pocock's mattpocock/skills repository topped GitHub Trending with 233,815 stars—more than double the second-place OpenAI Codex repo (115,131 stars). The repository contains no TypeScript code; instead it packages a senior engineer's workflows into markdown-based Skills with decision trees, checklists, and explicit constraints, organized into semantic directories such as /skills/typescript, /skills/testing, /skills/architecture, and /skills/debug. Notable Skills include /grill-me (adversarial interviewing), /tdd (red-green-refactor vertical slices), /caveman (~75% token reduction), /triage, and /git-guardrails-claude-code. The post contrasts Pocock's Procedure-style Skills (manually triggered, context-light) with Obra's Ability-style Superpowers (model-invoked, always in context), and highlights ptmrio/harness-subagent, a community tool enabling cross-framework orchestration between Claude Code, Codex, Cursor, and Grok Build under the principle that 'another Harness is not an oracle.' Together, these projects signal a shift in AI coding competition from raw model capability toward orchestration layers, context compression, and team-level Skill assets that could evolve into an npm-like Skill Registry.

Overview

On August 24, mattpocock/skills topped GitHub Trending with 233,815 stars—more than double second-place OpenAI Codex (115,131 stars), and well ahead of multi-model aggregators (47,945) and Apache Maka (2,342). The repo publishes Matt Pocock's daily-use Claude Skills from his .agents/ directory. It contains no TypeScript code—instead, it is a behavioral specification library that writes a senior engineer's judgment branches, checklists, and forbidden actions as markdown protocols consumable by Claude Code, Codex, and Cursor.

What's in the repo

Skills are organized semantically: /skills/typescript, /skills/testing, /skills/architecture, /skills/debug, /skills/planning, /skills/governance, /skills/meta. Each Skill is an engineering workflow with decision trees, checklists, error handling, and constraints—not a vibe prompt. Highlights:

/grill-me — 4–5 sentences that turn the Agent into an adversarial interviewer, pushing an idea until consensus
/caveman — cuts token consumption to ~75% reduction
/triage — turns issue handling into a state machine
/git-guardrails-claude-code — barriers before destructive git operations
/tdd — enforces red-green-refactor vertical slicing
/write-a-skill — a meta-Skill teaching how to write your own

The design metaphor: Skills are composable workflow atoms, not black-box prompts. Teams can version, audit, and roll out internal Skills as engineering assets—potentially evolving into an npm-like Skill Registry.

Procedure vs. Ability

Two Skill philosophies are emerging:

Procedure (Pocock) — Skills are manually/command-triggered workflows; descriptions don't enter the context window, so you never pay tokens for unused Skills. As Pocock puts it: *"I prefer to be in control. I don't want to delegate my thinking to the model."*
Ability (Obra's Superpowers, 270K stars) — the model decides when to invoke; descriptions persist in context. Installing 100 ability Skills means 100 descriptions leaking into every turn.

The underlying divide: human-led vs. model-led agency. A Chinese-community compromise, MageByte-Zero/spec-superflow, hard-codes rules at the IDE layer to bypass the Agent-as-its-own-judge blind spot.

Cross-Harness orchestration

Also on August 24, ptmrio/harness-subagent shipped cross-Harness orchestration:

Developers stay in their primary Harness (Claude Code, Cursor Agent, Grok Bot) and dispatch subtasks as one-shot subagents to other Harnesses (Codex, Grok Build, GPT, etc.)
Implemented via Claude Code headless mode, Codex exec, Grok Build CLI; compatible with the Agent Skills spec
Core idea: "another Harness is not an oracle"—having a model review its own work only reproduces its blind spots; cross-vendor verification catches single-model systemic defects

This contrasts with vendor-centric moves (OpenAI open-sourcing Codex Harness on 8/22; DeepSeek Harness rc.8 on 8/20 absorbing Codex and Claude Code as Profile Bundle subagents). Meanwhile Hermes burned 2T tokens in a single day with per-request contexts of 50K–100K tokens—pushing competition toward context compression, on-demand loading, and reasoning scheduling.

Four shifts defining the AI coding battlefield

1. Prompt Engineering → Skill Engineering: prompts are static text; Skills are executable protocols with decision trees, checklists, and state machines 2. Single framework → cross-Harness: Claude Code + Codex + Cursor + Grok Build assembled on demand 3. Model capability → scheduling capability: with frontier models tightly matched on benchmarks, the Harness orchestration layer decides outcomes 4. Individual prompt tricks → team Skill assets: private Skill Registries may become a more durable moat than model weights

Practitioner checklist

Install mattpocock/skills; start with /grill-me, /tdd, /caveman, /git-guardrails-claude-code
Evaluate a private Skill Registry, beginning with code review and security audit
Avoid vendor lock-in; require headless mode support (OpenCode, Codex exec, Grok Build CLI)
Pair high-throughput workloads with context compression and on-demand loading
Watch Skill-ecosystem versioning: akillness/oh-my-skills already integrates 15 of these Skills into a 113-skill unified package

Sources

AI Daily Report 2026-08-24 (Smartotics, GitHub Trending) + mattpocock/skills (233,815 stars)
ptmrio/harness-subagent repository and documentation
ai 技术日报 2026-08-24 (博客园); 腾讯新闻 Harness engineering updates 8/24 (Hermes 2T tokens/day, EnvHarness paper)
尧图网络, HQWC, 53AI, 爱客知识 coverage; akillness/oh-my-skills changelog