Skills System¶
TL;DR
Skills are SKILL.md files that act as SOPs for Claude Code. Type /morning and Claude follows a step-by-step procedure. Fifteen published skills cover daily planning, meeting prep, document review, content creation, task management, and weekly maintenance. Write your own in under 10 minutes.
What¶
A skill is a SKILL.md file that acts as a standard operating procedure (SOP) for Claude Code. When you type /morning, Claude reads the corresponding SKILL.md and follows its instructions step by step. Skills are versionable, inspectable, and modular - plain markdown files checked into git.
Why¶
System prompts are invisible and monolithic. SKILL.md files fix both problems. You can read them, edit them, diff them, and share them. Each skill owns one workflow. When a skill breaks, you fix one file. When you want a new capability, you write one file.
The alternative - cramming everything into CLAUDE.md - produces a bloated instruction file that Claude struggles to follow consistently. Skills decompose that complexity into focused, testable units.
How¶
SKILL.md Format¶
Every skill starts with YAML frontmatter:
---
name: morning
description: Generate daily plan with AI prioritisation. Run every morning (5 min).
argument-hint:
allowed-tools: Read, Write, Bash, Glob
model: sonnet
---
Key fields:
- name - The slash command trigger
- description - What it does and how long it takes
- allowed-tools - Which Claude Code tools the skill can use (security boundary)
- model - Which model to use (sonnet for routine work, opus for strategic analysis)
Below the frontmatter, the skill body defines the procedure: what to read, how to process, what to output.
The Gather-Analyse-Synthesise Pattern¶
Most skills follow the same three-phase structure:
sequenceDiagram
actor User
participant Skill as SKILL.md
participant Vault as Vault Files
participant Output as Output File
User->>Skill: /morning
activate Skill
Note over Skill: Phase 1: Gather
par Read in parallel
Skill->>Vault: tasks.md
Skill->>Vault: MEMORY.md
Skill->>Vault: theme/status.md
Skill->>Vault: calibration-log.md
end
Note over Skill: Phase 2: Analyse
Skill->>Skill: Apply leverage scoring
Skill->>Skill: Weight by strategic priority
Skill->>Skill: Draft follow-ups for stale items
Note over Skill: Phase 3: Synthesise
Skill->>Output: Write daily-plan.md
Skill->>User: Present prioritised plan
deactivate Skill
- Gather context - Read vault files in parallel (tasks, theme status, memory)
- Analyse - Apply domain logic (prioritisation, scoring, pattern matching)
- Output - Write results to a specific file or present to the user
Skills can spawn Task subagents for parallel execution. The /morning skill, for example, reads tasks and four theme status files simultaneously before synthesising priorities.
Published Skills¶
| Skill | Purpose | Runtime |
|---|---|---|
/morning |
Daily prioritised plan | ~5 min |
/brief |
Fast pre-meeting one-pager | ~2 min |
/challenge |
Red team / critical review | ~5 min |
/stress-test |
Multi-persona adversarial debate | ~15 min |
/transform |
Process content into vault | ~5 min |
/draft |
Outbound content (emails, LinkedIn) | ~5 min |
/prep |
Vault research before writing | ~10 min |
/prompt |
Structure thinking into briefs | ~30 sec |
/capture |
Store rhetorical weapons | ~30 sec |
/inbox |
Process captures and route | ~5 min |
/show |
Filtered task views | ~30 sec |
/weekly |
Archive, audit, review | ~30 min |
/changelog |
Inspect iteration decisions | ~1 min |
/visualise |
Diagrams from vault notes | ~2 min |
/evening |
End-of-day reflection | ~5 min |
Creating Your Own Skill¶
- Create
.claude/skills/[name]/SKILL.md - Add YAML frontmatter with name, description, allowed-tools
- Write the procedure as numbered steps
- Reference vault paths explicitly (skills should be self-contained)
- Test by running
/[name]in Claude Code
Key Insight¶
Skills are SOPs for an AI employee. The same principle applies: if you wouldn't trust a new hire to figure out a workflow from scratch every time, write it down. The SKILL.md file is the written procedure.
Customisation Points¶
- Model selection per skill - use cheaper models for routine work
- Tool restrictions - limit what each skill can access
- Argument passing - skills can accept parameters (e.g.,
/brief #project-a @alex) - Nesting - skills can reference other skills' outputs
Related¶
- System Overview - Where skills sit in the seven-layer architecture
- Three-Mode Search - Skills use parallel search during the gather phase
- Six-File Memory - Skills read memory files for context
- Self-Improvement Loop - Sessions generate improvement suggestions that refine skill behaviour