Company Skills Library — Building and Maintaining a Team-Wide Agent Skills Repository
Company Skills Library — Building and Maintaining a Team-Wide Agent Skills Repository
A shared repository where everyone writes down their whole team's complete workflow — and everyone gets the update. The key insight: downloading skills through a UI every time there's an update kills workflow; instead, set up auto-scan or hook agents to always pull the latest version.
What the pattern is
Shannon Holmberg (@shannholmberg) posted a thread on X laying out how to create and maintain a company-wide skills library. The core idea: instead of treating skills as something each person downloads individually from a UI (which leads to version drift and nobody knowing about improvements), treat a GitHub repo as the source of truth ("the Brain") and auto-pull the latest versions into every agent session.
The workflow
Bad path:
- Each person downloads skills through a UI every time
- Versions drift per person
- Nobody knows about improvements
- Skills rot
Good path:
BRAIN → AUTO-PULL → USE + LOG → PATCH → BRAIN
- BRAIN — a GitHub repo as the single source of truth for all team skills
- AUTO-PULL — at session start, the agent pulls the latest version of the company-brain skills
- USE + LOG — agents work on the real task, logging what worked and what failed
- PATCH — when someone improves a skill (observed mistake → noticed → patched), they push to the company-brain
- BRAIN — everybody gets the improvement on the next agent interaction
The improvement loop
The thread describes a tight feedback loop:
- Observe — someone makes a mistake
- Notice — pull the skill, edit it
- Patch — push to company-brain
And the cultural layer:
- Git / agent-friendly — use tools that work well with this pattern
- Use other people's skills — don't rebuild what exists
- Share what you learn — feed the loop
- Robot agents — automate the scanning and pulling
Why it matters
This is the team/organizational version of what individual developers are already doing with Claude Code skills (.claude/skills/), OpenClaw skills, and similar. The Anthropic guide anthropic-skills-guide covers how to build skills; this thread covers how to operate them at team scale — the governance, distribution, and improvement mechanics that make a skills library actually useful rather than a folder that 10% of the company knows how to use.
The insight echoes what loop-engineering argues: design loops (automations, worktrees, skills, connectors, sub-agents + memory) rather than prompting agents turn-by-turn. A skills library with an auto-pull + patch loop is a designed loop.
Related
- anthropic-skills-guide — the 32-page Anthropic guide on building, testing, and distributing Claude skills
- skills-managers — skills managers and marketplaces
- openskills — open skills collections
- loop-engineering — designing loops for agents rather than prompting turn-by-turn
- autonomous-dogfooding-skill — autonomous testing skill as an example of a skill that runs itself
Source
Shannon Holmberg (@shannholmberg) — X thread on company-wide skills libraries, August 2026.