The Starter Pack is part of Floom's broader product: a skill library that syncs across every agent, auto-updated. We built this pack first so you could see exactly what "curated" means in practice.
Federico built Rocketlist (500 WAU, Angel-backed) before starting Floom. He uses Claude Code, Cursor, and Codex daily and got tired of re-installing the same skills from scratch on every machine. Floom is the infrastructure he wanted to exist. MIT-licensed, built in the open.
Adam builds the distribution and packaging layer that makes skills portable across agents. Background in developer tooling and open-source infrastructure. Believes the skill ecosystem is the next big unlock for AI productivity.
Curated skills give AI agents a +18.6pp lift on real benchmarks (SkillsBench, arXiv:2602.12670). But kitchen-sink installs hurt: loading too many irrelevant skills forces the agent to evaluate 200+ options per task, increasing both latency and error rate. The lift only shows up when skills are picked carefully and invoked properly.
The activation side matters just as much. The AGENTS.md pattern that Vercel documented reaches 100% skill invocation, versus 70% for default installs. Every skill in this pack ships with its own activation rule so the agent knows when to reach for it.
Every agent user today rebuilds the same skills per machine, per project, per agent. Floom is the missing distribution layer: publish once, sync everywhere. The Starter Pack is the curated entry point. The broader product keeps everything auto-updated across all your agents without you thinking about it.
Skills.sh indexes 91,035 skills. Most are untested, self-generated, or duplicates of each other. We filtered by three criteria: proven in production by a real team, invoked consistently (not buried in docs), and non-overlapping with other skills in the pack.
Most of the 63 come from teams that already ship and use them in production: mattpocock, anthropics, vercel-labs, garry-tan's gstack, obra's superpowers, and others. Each has a source repo you can audit. We add the activation companions, the cross-agent install layer, and the curation; we don't intermediate the skill code.
For the full methodology, including how profile tags work and why 2-3 matched skills per task outperforms loading everything, see the docs.