Version-controlled manager for AI agent skills/prompt files
Problem
Developers working with coding agents (Claude Code, etc.) accumulate skill files, AGENTS.md instructions and prompt fragments scattered across repos, dotfiles and ad-hoc GitHub repos. There is no way to discover good community skills, keep them organized across projects and agents, verify they still work after model updates, or distribute company-wide skill sets to a team.
Opportunity
A 'package manager + registry' for agent skills: discover and install curated skills, sync them across projects/tools with versioning, run automated tests to check skills still work, and share team-managed skill sets with one command.
Market analysis
Discovery and installation are already claimed: Vercel's open 'npx skills' CLI positions itself as the package manager for agent skills, and multiple registries (ClaudSkills claims 191k+ crawled SKILL.md files) are live. The genuinely unsolved pieces are regression-testing skills against model updates and team-wide distribution with versioning — real problems, but ones an agent vendor can absorb natively at any time.
Market · Agent-heavy developers and platform teams; the Ask HN thread (315 points, 122 comments) confirms strong pain, but it is pain around a fast-moving ecosystem whose vendors want to own the workflow.
Pricing · Current tooling is free/open-source across the board; the only credible revenue is a team tier for managed skill sets and CI regression testing, comparable to other dev-team SaaS.
Pros
- + Verified strong demand signal from the HN thread.
- + Automated skill regression testing after model updates is genuinely unsolved and valuable to teams.
- + Team-managed, versioned skill sets with one-command install is an unclaimed enterprise wedge.
Cons
- − Vercel's 'npx skills' CLI plus skills.sh already own discover+install mindshare.
- − Multiple registries (ClaudSkills, Claude Marketplaces) are live and free.
- − Anthropic or any agent vendor can ship native skill management and erase the category.
- − Developer-tool willingness to pay for solo utilities is historically low.
Existing / similar tools
Source
Hacker News (Ask HN)
If you build here, do not build another registry — that race is over. The two durable wedges are verification and governance: a CI service that replays each skill against new model versions and flags behavioral drift, and a team layer that pins approved, versioned skill sets across agents. Both are boring, both are sticky, and both monetize per-seat instead of competing with free directories. The risk to price in is structural: if Anthropic ships first-party skill versioning, only the cross-vendor neutrality of your tool keeps it alive.