A skill is a plain SKILL.md in your repository: how your team does migrations, what your review bar is, the deploy dance nobody remembers. Agents load one when the work calls for it — and when an agent works something hard out, it proposes a new skill as a pull request you review.
You want the AI to stop making the same mistake. You do not want it quietly rewriting its own instructions in a system you cannot read, cannot diff and cannot roll back. Most “memory” features are exactly that — an opaque store that changes behaviour with no review step.
You explain the deploy order again. And again. It was right there in a thread from last month.
One person knows why the migration has to run twice. Nothing writes it down, so nobody else ever will.
If an agent can change what it believes without a diff, you cannot audit it and you cannot undo it.
A SKILL.md under `.decisive/skills/<name>/` with a one-line description at the top. An admin syncs it in from Settings → AI and switches it on.
An agent sees only the one-line description until it judges the skill applies, then pulls the full text in. A shelf of skills costs almost nothing until one is used.
Worked something out that will come up again? It opens a pull request plus a task to review it — never a live skill.
Merge the PR, the sync brings it in switched OFF, and an admin still has to enable it. Rejecting is just closing the pull request.
Nothing about a skill runs, and nothing about a skill reaches the network. It is instructions an agent reads.
By default every agent may load every enabled skill; an admin can narrow a particular agent to a subset.
A coding agent connected over MCP can propose a skill the same way. Free on every plan.
Free for up to 10 people, with 1,000 AI credits included. No credit card, no sales call, no onboarding deck.