When a sequence of tool calls achieves something useful on your site, that knowledge is worth keeping. Talos compiles successful runs into draft skills, scoped to your install, which you review before they become reusable.
A draft is often created automatically from a successful run, but it does nothing until you approve it. You can inspect, search, approve or reject each one, and every skill is exportable as markdown with diffs between versions.
Skills are stored in your own install and describe how work gets done here — your plugin set, your content structure, your conventions. Nothing is shared between sites.
Approving and replaying a skill escalate the risk level, so both are gated. A skill cannot quietly become a way to bypass tool policy.
Skills and their versions are stored in durable tables, and replay runs compatibility checks first so a skill built against an older site state does not misfire.
Draft → approve or reject → search → replay
Durable skills and skill_versions tables
Markdown, with version diffs
Site-local. Never shared between installs
No. Learned skills are records stored in your own WordPress database, not model training. Whether your prompts are used for training by your chosen AI provider is governed by that provider’s terms and your account settings with them.
Yes. Skills can be inspected, rejected and exported as markdown. Because they are versioned, you can see how one changed over time rather than only its current state.
Not automatically — they are deliberately site-scoped, since a skill encodes assumptions about a specific install. Markdown export gives you a way to carry an approach across manually.
Spend less time clicking through admin and more time moving your site forward.
