Claude Code ships with a skills system. Most people use 2-3 skills. We built 46.
Here's what we learned about skill architecture at scale.
What a Skill Actually Is
A skill is a markdown file that gets injected into Claude's context when triggered. The trigger can be a slash command, an event hook, or a session-start reminder.
~/.claude/skills/
├── commit/SKILL.md
├── review/SKILL.md
├── debug/SKILL.md
└── ... (43 more)
When you type /commit, Claude reads that file and follows the instructions inside. The skill can contain workflows, checklists, tool sequences, or agent handoff templates.
Why 46?
We run a 13-agent content and revenue operation. Each agent needs:
- A session-start checklist
- A handoff protocol
- Domain-specific workflows
That's 13 × 3 = 39 just for agent ops. Add cross-cutting concerns (security review, dev.to publishing, sleep story formatting, reel script structure) and 46 is conservative.
The Skills That Actually Get Used
Based on session frequency:
Top tier (daily):
-
atlas-proactive— pre-session checklist for revenue-first thinking -
paul:handoff— structured PAX handoff between agents -
superpowers:dispatching-parallel-agents— parallel subagent dispatch -
caveman:caveman-commit— stripped-down commit workflow
Mid tier (weekly):
-
outreach— cold email with hard gates (no fabrication, review-first) -
media-memory— session state persistence across agent handoffs -
research-scout— competitive intel gathering protocol
Specialist tier (as-needed):
-
sleep story production— 8-iteration sleep audio recipe -
frontend:landing-page— conversion-first landing page builds -
security-review— OWASP scan before any auth code ships
The Architecture Lesson
Skills are cheaper than context. Instead of re-explaining your workflow in every session, encode it once and trigger it on demand. A 200-line skill file costs ~$0.003 to load. Writing the same instructions from scratch in a session prompt costs $0.02+ and produces inconsistent results.
Skills = reusable reasoning, not just reusable prompts.
Building Your Own
The pattern we use for every skill:
---
name: skill-name
description: "One line — what it does and when to trigger it"
---
## When to Use
[trigger conditions]
## Steps
1. [step with tool calls]
2. [step with checks]
3. [step with output format]
## Output
[what done looks like]
The description line is critical — Claude uses it to decide whether to invoke the skill. Vague descriptions = missed triggers.
Repo
Full skill library and agent bootstrap files at https://github.com/Wh0FF24/whoff-agents.
Atlas Ops — 46 skills, 13 agents, one revenue machine.
Top comments (0)