SKILL.md won the format. It did not win the authoring.
Skills are becoming the unit of agent knowledge. The hard part is where they come from.
For about two years, getting something out of an agent meant writing a good prompt. Now half the field is talking about skills: small files of written-down competence an agent loads when it needs them. SKILL.md, Anthropic's open Agent Skills format, won that argument fast. It is portable, any compliant agent can load it, and it reads like a runbook instead of a prompt.
The format is the easy part. The hard part, the part almost nobody is solving, is where the skills come from and whether they are still true.
A hand-written skill is wrong by next week
Write a SKILL.md for how your team handles refunds and it is correct the day you write it. Then the threshold moves in a Slack thread, finance adds an exception over email, and your file still says last quarter's rule. The agent loads it and acts on a policy that is no longer true, with total confidence.
This is not a model problem. The model is fine. It is a knowledge problem. The real decision lives in the places work actually happens, not in the doc someone wrote once and never touched again.
Two things actually matter
First, provenance. Every rule an agent acts on should carry a citation to the exact message that set it, so the agent can verify the rule is current before it moves money or closes a ticket. A skill you cannot trace is a skill you cannot trust.
Second, freshness. The skill has to be derived from the live source, not hand-maintained. If editing the policy in Slack does not change the skill, you do not have a knowledge layer. You have a snapshot, and snapshots rot.
What we do about it
Lore reads the places your decisions live, Slack and Gmail and Notion, and auto-writes the SKILL.md with a citation on every line pointing back to the source. When the source changes, the skill re-emits, and the old rule is kept and marked superseded for audit. The agent always knows which version is current, and it can prove it.
The format was never the bottleneck. The authoring is.
We run it on ourselves
Our own company runs on three cited skills today: refund-handling, written from 47 real observations, incident-response from 41, and pricing-exceptions from 38. Every clause is clickable to the message where it was decided. We would rather show three skills you can audit than thirty you cannot.
We are early, in YC-review build, hand-onboarding the first design partners. But the bet feels more right every week. The next unlock for agents is not a smarter model. It is knowing how your company actually works, with the receipt attached.
If your agents are guessing at your own policies, we would love to talk. That is the whole reason we are building in the open.