Master AI Coding with Matt Pocock Skills: From Deep Alignment to Architecture in One Workflow

Overview

Matt Pocock Skills is a collection of lightweight, composable AI‑agent skills designed for Claude Code, Codex and similar coding assistants. The skills focus on deep alignment, shared language, feedback loops, deep modules and a global view to turn “vibe coding” into repeatable, high‑quality delivery.

Core Concepts

1. Deep Alignment (Grilling)

Before any code is written, the agent asks a series of focused questions to ensure it fully understands the plan. The command /grill-me runs the questionnaire; /grill-with-docs updates CONTEXT.md and creates ADRs while the interview proceeds.

“No one knows what they really want.” – David Thomas & Andrew Hunt

Typical workflow:

用户描述计划
↓
AI 逐个问题深度质询
↓
每个分支决策树被解决
↓
达成共享理解

2. Shared Language

Developers and domain experts use a common terminology file ( CONTEXT.md) and architecture decision records ( docs/adr/) to reduce communication overhead. Naming consistency cuts token usage by roughly 75 % (empirical claim, not a billing guarantee).

3. Feedback Loops

Combines test‑driven development ( /tdd) with systematic diagnosis ( /diagnose) in a six‑phase cycle: build feedback loop → reproduce → hypothesize → instrument → fix + regression test → clean‑up.

Phase 1: 构建反馈循环
↓
Phase 2: 复现
↓
Phase 3: 假设
↓
Phase 4: 仪器化
↓
Phase 5: 修复 + 回归测试
↓
Phase 6: 清理 + 事后分析

4. Deep Modules

Modules are kept small at the interface level but powerful underneath. The command /improve-codebase-architecture discovers shallow modules, proposes deeper replacements, and records the decision as an ADR.

5. Global View

/zoom-out

gives a high‑level map of a code region, useful when onboarding new code or planning refactors.

Installation & Quick Start

Run the one‑liner to add the skill set, choose the desired skills, pick the target agent and initialise:

# 1. 运行安装脚本
npx skills@latest add mattpocock/skills

# 4. 在代理中运行初始化命令
/setup-matt-pocock-skills

Verify the installation with /help inside the agent.

Key Commands

/grill-me

– deep alignment questionnaire. /grill-with-docs – alignment plus automatic updates to CONTEXT.md and ADRs. /tdd – red‑green‑refactor test‑driven development. /diagnose – systematic debugging with a nine‑step harness hierarchy. /triage – issue classification workflow (bug, enhancement, needs‑triage, etc.). /to-prd – generate a product‑requirements document from the current conversation. /to-issues – split a plan into vertical‑slice issues. /zoom-out – produce a global perspective of unfamiliar code. /improve-codebase-architecture – run a periodic architecture review. /caveman – “brief mode” that strips filler words, reducing token usage by ~75 % while preserving technical accuracy. /write-a-skill – scaffold a new skill with the required directory layout and SKILL.md template.

Best Practices

Combine skills in the recommended order: initialise → deep alignment with documentation → generate PRD → split into issues → triage → TDD → diagnose → architecture review. Keep CONTEXT.md up‑to‑date whenever new terminology, clarified concepts or domain definitions appear. Create ADRs only when three conditions hold: high reversal cost, missing context would surprise readers, and a real trade‑off has been evaluated.

Token Optimization with Caveman Mode

Use /caveman for long debugging sessions, massive code reviews, repetitive tasks or when the token budget is tight. The mode removes articles, filler phrases and polite language while preserving code blocks and precise error messages. Expected token reduction is around 75 % (individual variance).

FAQ Highlights

Supported agents – primarily Claude Code, but any agent that can load SKILL.md works.

Updating a skill – re‑run the install script.

Customising a skill – edit SKILL.md, add reference files, or scaffold a new skill with /write-a-skill.

Disabling a skill – delete its directory.

Token impact – skills are lightweight; /caveman provides the main savings.

Multi‑module projects – use CONTEXT-MAP.md to map several contexts.

Release cadence – the repository has 58 commits as of 2026‑04‑29; regular updates are recommended.

Conclusion

Matt Pocock Skills turn AI‑assisted coding into a disciplined workflow that emphasizes understanding, documentation, testing and architecture. By applying deep alignment, shared language and feedback loops, teams can replace “vibe coding” with repeatable, high‑quality delivery.