How it works: Skills

Some commands here show a name in angle brackets until you fill it in. Add your names on the setup page, and every command on every page fills them in.

Skills

A skill is a recipe Claude follows: a folder with a SKILL.md file in it. The top of the file has a name and a one-line description; the rest is step-by-step instructions, and some skills carry scripts or reference files beside it.

The problems they solve

You explain the same routine every time. Without a skill, how you like your day planned or a draft checked lives in your head, so you retype it each session and skip a step half the time. A skill writes the routine down once, and Claude follows it the same way in every session.

Improvised work is shallow. Ask "check my resume" and Claude offers a few opinions and stops. The review-writing skill makes it run a checker script, then one fresh reviewer per writing rule, and report only once they come back clean. Most of a skill's value is turning "do a good job" into named steps with a finish line.

Standing instructions crowd out the work. You could paste every routine into CLAUDE.md, but that file is read in full every session, and every line takes room in the context window. A skill costs one line, its description, until it is needed; then Claude reads the whole recipe and follows it. That is why dozens of installed skills slow nothing down.

Corrections don't last. Tell Claude "no more than one fish dinner" in a chat and it holds for that chat. Put it in the skill and it holds for every run after.

Good routines stay with whoever wrote them. A skill is a folder, so it travels: Matt Pocock's skills are the workflows he uses day to day, and this pack is someone else's routines handed to you.

How a skill starts

A skill starts in one of two ways:

You type its name after a slash: /bod. Type / on its own to see the menu, and keep typing to filter it.
Claude decides the skill fits what you asked and starts it itself. "Remember this decision" starts decision-capture without you naming it.

Some skills switch the second way off, so they only run when you ask; those are the ones that take over a whole session, like teach.

Where they come from

Your skills come from three places:

~/.claude/skills/: the ones this pack installed, available everywhere.
Plugins: bundles of skills installed from a marketplace. Their names carry the plugin's name in front, such as /mattpocock-skills:teach. This pack installed Matt Pocock's skills and the Obsidian skills.
A folder's own .claude/skills/: skills that only exist when Claude runs there, like the two in the optional job-search workspace.

The pack's own skills, and what each one fixes:

Skill	What it does	The problem it solves
`/bod`	Tidies yesterday's capture and lists what is open	Starting the day without knowing what is waiting, while raw notes pile up
`decision-capture`	Records "chose X over Y because Z" to the vault	Months later you remember what you picked but not why, and argue it again
`meeting-capture`	Files meeting notes into person and project notes, and preps you for the next one	Notes stay in a chat or a notebook and never reach where you look before the next meeting
`/memory-groom`	The full tidy of the vault	Memory that piles up, repeats itself, and goes stale; see the brain, below
`/weekly-review`	Writes up the week from the vault	Weeks blur together, and a stalled project goes unnoticed
`review-writing` (optional job search)	Checks a draft against the writing rules, one fresh reviewer per rule	A draft Claude wrote can sound like Claude wrote it, and the writer misses its own tells
`verify-claims` (optional job search)	Checks each claim against your records, then has a stranger read it	A resume line you can't back up in an interview, or one an outsider can't follow

To look inside one, ask "show me how the bod skill works", or read it yourself with cat ~/.claude/skills/bod/SKILL.md.

Writing your own skills

Any routine you find yourself explaining to Claude twice can become a skill. The quickest way is to ask:

Ask Claude

Make me a skill that turns a grocery list into a week of dinners.

Claude writes the file, and you try it. Here is roughly what it produces, saved as ~/.claude/skills/meal-plan/SKILL.md:

File

---
name: meal-plan
description: Turn a grocery list into a week of dinners. Use when I paste a grocery list or ask for a meal plan.
---

1. Read the list. If I haven't said what is already in the kitchen, ask once.
2. Plan seven dinners, using the food that spoils first early in the week.
3. Save the plan to ~/notes/meal-plan-<today's date>.md: one line per day, then a
   shopping list of what is missing. Done when every ingredient on the plan is either
   on my list or on the shopping list.

What each part does:

The folder decides where the skill works. ~/.claude/skills/ makes it available in every session; .claude/skills/ inside a folder makes it exist only when Claude runs there.
name is what you type after the slash: /meal-plan.
description is the only line Claude sees until the skill runs, so it decides when Claude reaches for it. Say what it does and the situations it is for, in the words you actually use.
Add disable-model-invocation: true under the name for a skill only you should start. Claude then never starts it on its own, and its description stops taking up room in every session.
The steps are what Claude does, in order. A step works best when it ends on something Claude can check, like "done when every ingredient is accounted for" in step 3.

Then improve it the way you would with a person: run it, say what went wrong ("it planned fish twice; no more than one fish dinner"), and ask Claude to update the skill. If a new skill doesn't show up in the / menu, /exit and start Claude again. Give your skills their own names; the pack's updates replace its own skills (bod, memory-groom, and the rest), so keep your edits in skills of your own; a copy of an edited pack skill goes to ~/.config/leaven/backup-<stamp>/ first.

Writing for agents

A skill, a CLAUDE.md, and the rules in a folder are all writing for an AI to follow, and that has its own craft. Matt Pocock's writing-for-agents skill holds it, and Claude reads it on its own whenever it creates or edits a skill or a CLAUDE.md. Ask for it by name to review something you wrote:

Ask Claude

Use writing-for-agents to review my meal-plan skill.

The main ideas, in plain terms:

The description is a trigger. Write it for when the skill should run, and keep it short, because every description sits in every session.
Say what to do. "Write one-line answers" works better than "don't write long answers", because naming the unwanted thing puts it in Claude's head.
Give each step a finish line Claude can check. "Done when every receipt is filed" beats "file the receipts".
Keep the main file short. Long reference material, like a list of 50 recipes, goes in a second file beside SKILL.md, and the steps say when to read it.
Say each thing once. When two files give the same instruction, one of them goes stale.