Skills Patterns

Claude Code skill development patterns I’ve actually used. Every entry has a date and context.

2026-03: /deploy — Spoke Repo Deployment Skill

Problem: Deploying a spoke repo required 3 manual steps: commit, push, trigger Cloudflare rebuild. Repetitive and error-prone.

Context: Daily documentation workflow across 15 domus-* repos.

The Fix:

# Skill definition
name: deploy
description: Push spoke repo and trigger rebuild
tools:
  - Bash(git:*)
  - Bash(echo:*)
# Usage
/deploy

# What it does:
# 1. Validates repo name starts with domus-
# 2. Commits uncommitted changes (if message provided)
# 3. Pushes to GitHub origin
# 4. Triggers Cloudflare Pages rebuild via empty commit to domus-docs

Rule: Skills should encapsulate multi-step workflows that are repeated daily. Keep tool access minimal — /deploy only needs git and echo.


2026-03: /worklog — Daily Worklog Creation

Problem: Creating daily worklogs required creating the file, adding 8 partial includes, and updating navigation. Manual and inconsistent.

Context: Daily chronicle system, domus-captures.

The Fix:

# Skill definition
name: worklog
description: Create daily worklog from template
tools:
  - Bash(date:*)
  - Bash(mkdir:*)
  - Bash(ls:*)
  - Read
  - Write
# Usage
/worklog           # Creates today's worklog
/worklog tomorrow  # Creates tomorrow's
/worklog 2026-04-15  # Specific date

Rule: Skills that create files should parse flexible date arguments. Include date:*, mkdir:*, ls:* for filesystem operations. Template consistency > speed.


2026-03: Skill Design — Minimal Tool Access

Problem: Skills with broad permissions are security risks. A deployment skill shouldn’t be able to edit arbitrary files.

Context: Designing skill permissions for dots-quantum (stowed globally).

The Pattern:

Each skill gets ONLY the tools it needs:

  • /deploy: Bash(git:*), Bash(echo:*) — can push but can’t edit files

  • /worklog: Read, Write, Bash(date:*), Bash(mkdir:*), Bash(ls:*) — can create files but can’t run arbitrary commands

Rule: Principle of least privilege applies to AI tools. Over-permissioned skills are as dangerous as over-permissioned users.