Post

mattpocock/skills, Part 1: Get Started Without Memorizing 25 Skills

No need to memorize 25 skills: how to install mattpocock/skills, what setup-matt-pocock-skills adds to your project, and how ask-matt suggests your next step.

mattpocock/skills, Part 1: Get Started Without Memorizing 25 Skills

You’re about to add per-seat billing to an existing subscription service, but you’re not sure whether to start by clarifying requirements, writing a spec, adding tests, or just diving into the code. You’ve heard that mattpocock/skills can help an AI coding assistant handle this kind of work more methodically, but when you open the list, you’re greeted by 25 skill names. Do you really have to understand every one of them before you can begin?

You don’t. Your first time through takes just two steps:

  1. Use setup-matt-pocock-skills to record your project’s working rules.
  2. Describe your problem to ask-matt, and let it suggest the next skill to use.

As I see it, ask-matt is more of a help desk than a manager that runs the whole pipeline automatically. It only tells you which skill to use next; you still have to start that skill yourself.

mattpocock/skills SeriesPart 1

I based this post on version 1.2.3. The official plugin’s plugin.json includes 25 skills. The repository also contains 9 items that are still experimental and 4 that aren’t currently being promoted, listed in in-progress/README.md and misc/README.md respectively. This series only covers the 25 included in the official plugin.

On 2026-10-04, 1.3.0 and 1.3.1 took the official plugin from 25 skills to 27 (checked 2026-10-19):

  • resolving-merge-conflicts is gone.
  • implement-spec, pr, and retro are new, and ask-matt suggests them too.
  • The project glossary was renamed from CONTEXT.md to GLOSSARY.md. Skills in the new releases only look for the new name, so rename an existing file with git mv.

I plan to fully update this series for the new releases.

I think of a skill as a set of instructions that tells an AI coding assistant how to do a specific job. Examples include interviewing you about requirements or breaking a spec down into individual tasks. An AI coding assistant is a tool such as Claude Code or Codex that can read and write the files in your project and write code for you. mattpocock/skills is a complete set of skills published by Matt Pocock.

Installing and Invoking Are Separate Questions

The installation method determines where the skill files live and who is responsible for updating them. The invocation method determines who can tell a skill to start working.

A six-panel illustration of mattpocock/skills: choose the Claude Code plugin or skills.sh, run setup to create the project configuration, then let ask-matt suggest the next skill. A feature request moves through requirements clarification, spec writing, work breakdown, and implementation in order The full path for first-time use: choose an installation method, run setup, then let ask-matt suggest the next skill.

Pick One Installation Method

mattpocock/skills treats .agents/install-block.md as the single source of truth for installation instructions. Choose based on the tool you use:

Your situationInstallation methodMaintenance afterward
Using Claude Code, no plans to modify the skillsOfficial pluginUpdates automatically; contents can’t be modified
Using Claude Code, want to modify the skillsskills.shFiles are copied into your project; you modify and update them
Using Codex or another tool that supports skillsskills.shFiles are copied into your project; you modify and update them

I suggest picking only one of the two methods. If you use both, every skill shows up twice.

Claude Code is Anthropic’s AI coding assistant, while Codex comes from OpenAI. The command to install the official plugin for Claude Code is:

1
claude plugins install mattpocock-skills

You can also type this directly inside Claude Code:

1
/plugin install mattpocock-skills

If you want to modify the skills, or you use Codex or another tool, use the skills.sh installer instead:

1
npx skills@latest add mattpocock/skills

skills.sh lets you check off which skills to install and which AI coding assistants to install them for. Be sure to include setup-matt-pocock-skills, and include ask-matt too if you want to follow along with this post. Any other skill that ask-matt recommends later also has to be installed before you can use it.

With skills.sh, the command to update a single skill is:

1
npx skills@latest update <name>

Understand Who Can Invoke a Skill

mattpocock/skills sorts skills into two categories based on who can invoke them:

TermWho can invoke it
user-invokedOnly a human, by typing its name
model-invokedA human can invoke it, and the AI can also pick it up on its own when the task calls for it

A user-invoked skill can in turn call a model-invoked skill, but it cannot call another user-invoked skill. The full set of constraints is documented in the invocation rules.

Both setup-matt-pocock-skills and ask-matt are user-invoked, so you have to type their names yourself. Claude Code and Codex use different syntax:

ToolExample input
Claude Code/setup-matt-pocock-skills
Codex$setup-matt-pocock-skills

In Codex, you can type $ and then pick a skill from the list, or type $skill-name directly. You can verify this syntax in OpenAI’s Codex skills documentation. For other AI coding assistants, use their own skill menu or input syntax.

I think this constraint also explains why ask-matt can only make suggestions. For example, grill-with-docs, which it often recommends, is also user-invoked, so you have to type it yourself. It also won’t automatically go on to run to-spec, to-tickets, and implement.

Setup Inspects, Then Asks, Then Writes

Before a project first uses the skills that focus on software development, you need to run setup-matt-pocock-skills once. That’s because these skills (grouped under “engineering” in the repo) need some basic information about your project. Setup isn’t an installer that does the same thing every time; it inspects first, asks questions next, and writes files only at the end:

flowchart TD
  A[Inspect the project] --> B[Confirm issue tracker]
  B --> C{triage installed?}
  C -->|Yes| D[Confirm triage labels]
  C -->|No| E{Multiple subprojects?}
  D --> E
  E -->|No| F[One shared glossary and ADR set]
  E -->|Yes| G["Ask: one per subproject?"]
  F --> H[Show drafts to edit]
  G --> H
  H --> I[Write after you approve]

First, it looks at the current state of the project:

  • Which platform hosts the code, GitHub for example.
  • Whether the project already has AI instruction files (AGENTS.md or CLAUDE.md).
  • A glossary and decision records.
  • Configuration files left over from a previous run.
  • Whether the project is made up of multiple subprojects (a monorepo).

The glossary (CONTEXT.md) records definitions of project-specific terms, and decision records (ADRs) capture the reasoning behind important technical choices. I cover both in detail in the next post.

The issue tracker is the place where your team records to-do items. If your code is on GitHub, setup suggests GitHub Issues by default; if it’s on GitLab, it suggests GitLab Issues.

You can also track work in Markdown files inside the project folder. If you use another tool such as Jira or Linear, you describe how to work with it in a short paragraph.

The triage labels are used by triage, which sorts incoming external reports, so setup only asks about them if triage is installed.

After you confirm, it modifies or creates the following files:

LocationWhat gets writtenCondition
CLAUDE.md or AGENTS.mdAn ## Agent skills section pointing to the configuration files belowEdits CLAUDE.md if it exists, otherwise AGENTS.md
docs/agents/issue-tracker.mdWhere the issue tracker is and how to use itAlways created or updated
docs/agents/domain.mdHow the glossary and decision records are organizedAlways created or updated
docs/agents/triage-labels.mdThe names of the triage labelsCreated or updated only if triage is installed

If the project has neither instruction file, it asks which one you want to create rather than deciding on its own. If a file already contains an ## Agent skills section, it updates that section in place instead of adding a duplicate.

Setup only writes the label names into docs/agents/triage-labels.md; the process as written has no step that creates those labels on GitHub or GitLab. So before you start triaging, I’d first make sure the label names in your issue tracker match the configuration file.

Before writing anything, it must show you drafts of the ## Agent skills section and each configuration file so you can edit and approve them. Once written, these files are just ordinary project documents that you can edit directly. You only need to run setup again if you want to switch issue trackers or start over from scratch.

ask-matt Suggests a Starting Point for Your Situation

ask-matt doesn’t expect you to memorize every skill first. Just describe your current situation, and it suggests which skill to start with. Common starting points:

Current situationSuggested starting point
In a project, want to talk through a half-formed ideagrill-with-docs
No project folder, just want to talk an idea throughgrill-me
Need to sort through bug reports and requests submitted by otherstriage
Hitting a bug that’s hard to reproduce, intermittent, or recently introduceddiagnosing-bugs
The scope is fuzzy and won’t fit in a single conversationwayfinder

The main path for typical feature development starts with grill-with-docs, which clarifies requirements through an interview. Some questions can only be answered by building something you can actually try out, such as what a screen should look like. Then the flow takes a temporary detour:

flowchart TD
  A[grill-with-docs interview] --> B{Need something to try?}
  B -->|No| F[Requirements clear, continue]
  B -->|Yes| C[handoff writes progress]
  C --> D[prototype in new conversation]
  D --> E[handoff brings result back]
  E --> A

prototype builds a throwaway prototype.

If you’re sure the work will take several conversations to complete, use to-spec next to put together a spec and to-tickets to split the spec into individual tickets. Finally, open a fresh conversation for each ticket and implement it with implement.

Internally, implement uses tdd (a development approach where you write the tests first, then the code) to build the feature, and runs code-review before committing the changes.

Everything from the requirements interview through to-tickets should stay in the same conversation, so that later steps build on the earlier discussion. Only after the tickets are created does each one get its own new conversation. That works because each ticket already spells out the information it needs, so it doesn’t depend on earlier conversation history. The full rules are in ask-matt/SKILL.md.

On the left, a progress bar shows someone memorizing 25 skills; on the right, ask-matt asks which step they're stuck on Memory test canceled. Time to get to work.

Walkthrough: Adding Per-Seat Billing to an Existing Service

I reasoned through this section from the rules as written in version 1.2.3. Assume the project is made up of several subprojects, the code is hosted on GitHub, and work is tracked in GitHub Issues.

Using Codex as the example, first install the skills you need with skills.sh. Here I’ll assume every skill mentioned below has been selected. Then type:

1
$setup-matt-pocock-skills

Setup notices the code is on GitHub and suggests sticking with GitHub Issues. It also sees from the folder structure that the project consists of multiple subprojects, so it asks whether the glossary should be one shared copy or one per subproject.

If triage is installed as well, setup also asks whether to keep the five default labels:

1
2
3
4
5
needs-triage
needs-info
ready-for-agent
ready-for-human
wontfix

In order, these five labels mean “awaiting evaluation,” “waiting for more information,” “ready to hand off to the AI,” “needs a human,” and “decided not to fix.” If your team’s issue tracker already uses different label names, you can map them to those names instead to avoid ending up with duplicate labels.

Once you confirm these choices, the AI coding assistant first shows you what it’s going to write. Only after you approve does it modify CLAUDE.md or AGENTS.md and create the configuration files under docs/agents/.

With setup done, describe the work you’re stuck on:

1
$ask-matt We need to add per-seat billing to an existing subscription system. Right now there's only a short requirements description. The requirements should be clear after one conversation, but the implementation will happen over several sessions.

Following its rules, ask-matt will suggest starting with grill-with-docs. That skill works through a few questions: how pricing is calculated, when a seat starts and stops being used, the billing cycle, and how existing customers are migrated. The flow doesn’t continue automatically; you have to type the next skill yourself:

1
$grill-with-docs

Once the requirements are clear, stay in the same conversation and type the following in order:

1
2
$to-spec
$to-tickets

to-tickets splits the spec into smaller tickets and marks which ones must be done first and which have to wait until earlier ones are finished.

After the tickets are created, open a new conversation for each ticket and type:

1
$implement <ticket URL or number>

This workflow only handles the order of the work. It won’t decide your pricing rules, how seats are counted, or how old data is migrated for your team. Those product and technical decisions still have to be made by the team during requirements clarification.

How to Choose on Your First Run

flowchart TD
  A{One-off, outside a project?} -->|Yes| B[Skip full project setup]
  A -->|No| C[Pick one install method]
  C --> D[Run setup once]
  D --> E{Know the next skill?}
  E -->|Yes| F[Skip ask-matt, invoke it]
  E -->|No| G[Ask ask-matt]

Even when you know the next skill, skills that read your issue tracker or glossary configuration still need setup to be completed first.

In this post I checked the design and usage of version 1.2.3; that doesn’t mean this workflow will necessarily shorten development time. I think its direct benefits are that it turns your project conventions into written documentation and suggests a suitable skill when you’re unsure what to do next. The next post covers how to clarify requirements and build a vocabulary your team shares.

Further Reading

This post is licensed under CC BY 4.0 by the author.