# Skills For Real Engineering
**作者**: AlphaSignal AI
**日期**: 2026-04-26T19:54:04.000Z
**来源**: [https://x.com/AlphaSignalAI/status/2048809516993556845](https://x.com/AlphaSignalAI/status/2048809516993556845)
---
## A 21-skill prompt collection that encodes a real Claude Code workflow, what to install first and how.
mattpocock/skills crossed 28,000 GitHub stars on April 27, 2026, up from 18,700 two days earlier!
The repo contains no application code. The whole project is a folder of markdown files.
Who it's for: Engineers running Claude Code (or similar agents) who want a battle-tested workflow stack instead of a kitchen-sink CLAUDE.md.
> **Matt Pocock@mattpocockuk**: [原文链接](https://x.com/mattpocockuk/status/2048490818848075846)
>
> Nearly 23K stars for a collection of markdown files I wrote
> I guess they must be pretty good
> I want to invest more time in this repo. So, folks who starred it, what can I do to make these skills more obvious to you?
> - A docs site for the skills?
> - Send them to plugin
## Repo snapshot
## Context
A Reminder: Anthropic released the Agent Skills spec as an open standard on December 18, 2025. The skill contains a SKILL.md file with YAML frontmatter (name, description) plus optional reference docs and scripts. The agent loads only the name and description at startup. The full SKILL.md loads on demand. The pattern is called progressive disclosure.
@mattpocockuk pushed his personal ~/.claude/skills/ directory to GitHub on February 3, 2026. The README tagline: "My agent skills that I use every day to do real engineering, not vibe coding."
## What's actually inside
The README organizes skills by category. The five he names as his daily chain are the spine.
## The 5-skill daily chain
/grill-me (Check "current limitations" section before running it)
The whole file is 11 lines including YAML frontmatter. The agent walks the design tree, asks one question at a time, recommends an answer for each, and explores the codebase when a question can be answered from code instead of from the user. The frontmatter description is longer than any single paragraph in the body.
/to-prd
Synthesizes the current conversation into a Problem / Solution / User Stories / Implementation Decisions / Testing Decisions / Out-of-Scope document and files it as a GitHub issue via gh issue create. It does not interview. It uses what is already in context.
/to-issues
Breaks a PRD into vertical-slice "tracer-bullet" issues. Each slice cuts through every layer (schema, API, UI, tests) end-to-end. The skill marks each slice HITL or AFK and files them in dependency order.
/tdd
Enforces a vertical-slice red-green-refactor loop. The skill explicitly bans the horizontal anti-pattern (write all tests, then all code), citing that bulk-written tests verify imagined behavior rather than actual behavior. Ships supporting docs: tests.md, mocking.md, deep-modules.md, interface-design.md, refactoring.md.
/improve-codebase-architecture.
It looks for "deepening opportunities" where shallow modules can collapse into deep ones. Ships a LANGUAGE.md file that fixes the architecture vocabulary and forbids synonyms. Only Module, Interface, Implementation, Depth, Seam, Adapter, Leverage, and Locality are allowed. Words like "component," "service," "API," and "boundary" are banned inside the skill's own output.
The other 11 README skills
## Why it's different
The format is not the differentiator. Anthropic shipped the SKILL.md spec four months ago. The interesting decisions are in what Pocock encoded and what he refused to encode.
Process encoding, not knowledge dumping. grill-me does not tell the agent what to think. It encodes how to think: walk the tree, one question at a time, recommend an answer, explore the codebase when possible. Same pattern across tdd, to-prd, and improve-codebase-architecture. Any modern model already knows what TDD is. The skill enforces the discipline.
Narrow scope per skill. No "superpowers" catchall. Each skill targets one workflow with explicit "Use when…" triggers in its YAML description. The agent picks the right skill the way a developer picks the right Unix tool. The longest skill in the repo is github-triage at roughly 170 lines. grill-me is 11.
Intentional auto-invocation control. The disable-model-invocation: true flag is a structural decision most skill authors miss. He distinguishes between skills the agent should reach for autonomously (when the user mentions TDD, when they file a bug) and skills the user should reach for manually (when they want to map an unfamiliar codebase). Most skill collections leave that line undrawn and let the agent guess.
Personal artifact, not a packaged framework. The collection was exported straight from his ~/.claude/skills/ directory without rebranding or abstraction. The category tree (Planning & Design, Development, Tooling & Setup, Writing & Knowledge) maps to one engineer's daily workflow, not to a generic AI-dev-kit taxonomy. Reading any single SKILL.md surfaces the actual decision Pocock made for his own use, not a sanitized recommendation written to teach.
## How to use it
Install
Each skill is installed individually. Exact commands from the README:
```
npx skills@latest add mattpocock/skills/grill-me
npx skills@latest add mattpocock/skills/to-prd
npx skills@latest add mattpocock/skills/to-issues
npx skills@latest add mattpocock/skills/tdd
npx skills@latest add mattpocock/skills/improve-codebase-architecture
```
The CLI is from vercel-labs/skills. By default it drops each skill into .claude/skills// (project scope). Add --global to install at user scope (~/.claude/skills//).
The workflow walkthrough
1. Idea phase. Type /grill-me. The agent asks one question. The user answers. The agent walks the design tree until it stops asking and proposes the plan.
2. PRD phase. Type /to-prd. The agent synthesizes the grill-me transcript into a structured PRD and files it as a GitHub issue via gh issue create. Whatever was in context becomes the PRD.
3. Slicing phase. Type /to-issues against that PRD. The agent breaks it into independently grabbable slices, marks each HITL or AFK, and files them in dependency order so "Blocked by" links resolve correctly.
4. Implementation phase. Pick an AFK slice. Type /tdd. The agent writes one failing test, gets confirmation on the public interface, writes minimal code to pass, and repeats. No bulk test files.
5. Architecture phase. Type /improve-codebase-architecture. The agent reads CONTEXT.md and docs/adr/ if present, applies the LANGUAGE.md vocabulary, surfaces deepening opportunities, and drops into a grilling conversation on whichever candidate the user picks.
Backlog assumption. to-prd, to-issues, triage-issue, qa, and github-triage all hardcode gh issue create. Linear, Jira, and Beads users currently fork. He named multi-backlog support as a roadmap item on April 27.
## Multi-agent compatibility
The SKILL.md spec is portable across every coding agent that adopted Anthropic's standard. The repo follows it strictly, so the same folders run on Claude Code, Codex, OpenClaw, and Hermes Agent without modification.
The npx skills CLI defaults to the Claude Code path. For Codex, OpenClaw, or Hermes installs, manual copy of the SKILL.md folder into the target agent's skills directory may currently be required.
## Current Limitations
Auto Mode overrides skill instructions. April 26, Pocock flagged that Claude Code's Auto Mode silently injects system-prompt instructions pushing the agent toward autonomous execution, which can override skills like grill-me that pause for user input. Running Opus 4.7 in Auto Mode, he observed the agent skipping the grill interview entirely and jumping straight to implementation. He called the behavior "messing with all my skills" and "the cost of not owning the whole flow.
GitHub-only backlog assumption. to-prd, to-issues, triage-issue, qa, and github-triage hardcode gh issue create. Linear, Jira, and Beads users have to fork the skill body or skip it. He named multi-backlog support as a roadmap item on April 27.
No documentation site. Discovery today is the README plus the "Use when…" descriptions in each SKILL.md frontmatter. He confirmed April 27 that a docs site is on the roadmap. Not live as of publication.
No auto-updater. Reinstall via npx skills@latest add is the only way to pull updates. Version pinning happens at the commit-SHA level inside the vercel-labs/skills CLI, not at the user level.
Two skills are not portable. obsidian-vault hardcodes a personal Obsidian path. scaffold-exercises depends on pnpm ai-hero-cli internal lint, specific to Pocock's AI Hero course infrastructure. Both require editing before reuse.
No skill-level evals. He has not shipped tests or evaluation runs against the skills themselves. His own meta-skill, write-a-skill, advises evals as part of authoring but does not implement them for the repo.
So the best recommendation is to install the five-skill daily chain today, run /grill-me with Auto Mode off until the override behavior settles, fork the issue-filing skills if the team is not on GitHub, and watch for the docs site in May.
## Who benefits and who doesn't
Who benefits: engineers running Claude Code on production codebases who want repeatable design reviews, TDD discipline, and structured PRD-to-issue workflows. Teams already on GitHub Issues that want feature-planning and bug-triage conversations converted into durable artifacts. Skill authors looking for a reference implementation of narrow-scope skill design.
Who should skip: teams on Linear, Jira, or Beads until multi-backlog support ships. Engineers using agents that have not adopted the SKILL.md spec. Anyone looking for "do everything" autonomous coding skills, since Pocock's repo is deliberately the opposite.
## Practitioner implication
Engineers using Claude Code can now adopt this skill chain in five npx commands, now that the SKILL.md standard is portable across every major coding agent.
## AlphaSignal Take
The growing 28,000-star count is misleading as a signal of novelty. The repo is not novel infrastructure. It is curated workflow encoding on top of an existing four-month-old open standard. The structural surprise of 2026 is that taste and narrow scope travel further than feature breadth or framework reach.
The lesson for skill authors is not to copy Pocock's skills. The lesson is to copy his discipline: write an 11-line skill that does one thing, then stop.
## Links
- github.com/mattpocock/skills
Follow @AlphaSignalAI for more content like this.
Check out AlphaSignal.ai to get a daily summary of top models, repos, and papers in AI. Read by 280,000+ devs.
## Questions?
What is mattpocock/skills? A 21-skill open-source collection of SKILL.md-formatted prompts authored by Matt Pocock for Claude Code and other compatible AI coding agents. It encodes process workflows including TDD, design grilling, PRD writing, vertical-slice issue breakdown, and codebase architecture review. MIT licensed, released February 3, 2026.
How do you install Matt Pocock's skills? Run npx skills@latest add mattpocock/skills/ for each skill. The CLI is from vercel-labs/skills and drops the skill folder into .claude/skills/ (project scope) or ~/.claude/skills/ (user scope, with --global).
Which skills should engineers install first? The five-skill daily chain Pocock himself names as his daily workflow: grill-me, write-a-prd, prd-to-issues, tdd, and improve-codebase-architecture. They compose into a complete idea-to-merged-PR workflow.
Does mattpocock/skills work with Codex, OpenClaw, or Hermes Agent? Yes. The skills follow the open SKILL.md spec released by Anthropic in December 2025, which Codex, OpenClaw, and Hermes Agent all adopted. The npx skills CLI installs to the Claude Code path by default, so manual copy into the target agent's skills directory may currently be required.
What does the grill-me skill actually do? It instructs the agent to interview the user one question at a time, walk every branch of the design decision tree, recommend an answer for each branch, and explore the codebase instead of asking when the answer can be derived from code. The whole skill is 11 lines including YAML frontmatter.
## 相关链接
- [AlphaSignal AI](https://x.com/AlphaSignalAI)
- [@AlphaSignalAI](https://x.com/AlphaSignalAI)
- [Apr 27](https://x.com/mattpocockuk/status/2048490818848075846)
- [174K](https://x.com/mattpocockuk/status/2048490818848075846/analytics)
- [@mattpocockuk](https://x.com/@mattpocockuk)
- [flagged](https://x.com/mattpocockuk/status/2048387101981323409)
- [github.com/mattpocock/skills](https://github.com/mattpocock/skills)
- [@AlphaSignalAI](https://x.com/@AlphaSignalAI)
- [AlphaSignal.ai](http://alphasignal.ai/)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [1:00 AM · Apr 28, 2026](https://x.com/AlphaSignalAI/status/2048809516993556845)
- [8,031 Views](https://x.com/AlphaSignalAI/status/2048809516993556845/analytics)
---
*导出时间: 2026/4/28 15:25:48*