How curation works
How curation works
While you work, your coding agent has the team's skills in front of it and standing instructions to capture what a session earns: after a complex task, a tricky error, or a non-obvious workflow, it writes or patches a skill and publishes it right in the session, where you can watch the checks run.
Roughly every five turns comes a checkpoint: the agent looks back over the recent work and asks whether anything worth keeping slipped past in the flow. On Claude Code the checkpoint is enforced mechanically; on agents with no such channel it is written into their instructions, and every checkpoint announces its outcome in one line either way.
Before anything leaves your machine
Four deterministic checks run in plain code — no model involved:
- valid — is it a well-formed skill, actually written
- safe — does it carry anything credential-shaped (blocked, never sent)
- helps — does it plausibly improve the task it came from
- findable — would anyone locate it; near-duplicates are refused
Session transcripts never leave the machine they were typed on. Only finished, checked skills do.
On the hub
Every change is a new immutable version — nothing is edited in place, and any version can be restored. The hub re-runs the safety checks itself and verifies you edited the version you actually read, so two people can never silently overwrite each other. When two skills turn out to cover one subject, they are consolidated: the broadest becomes the umbrella, the sources are archived, recoverable.
On everyone else's machine
A new version arrives automatically and waits as a pending proposal until that person approves it. CLI merge requests are the central exception: the merged draft waits in the shared hub until an owner/admin approves it, because approval is what archives the source skills. Two local-activation exceptions: your own skills activate for you instantly, and locked company standards are always active everywhere, because that is what a standard is.