A compiled binary tells an agent nothing about how to use it — what the flags mean, the gotchas, the right recipe. That knowledge is the toolkit's skills: hand-authored org files an agent reads on demand.

The manifest is a need-keyed index; each skill is read only when relevant, so an agent spends context on the one tool it's using, not the whole catalog.

A good skill has a fixed shape: when to use this (and when not), the mental model (why before how), a workflow of runnable steps, common pitfalls found empirically, and a verification checklist. Breadth is the common surface plus the gotchas — not an exhaustive flag dump.

#+begin_src org

Trigger + a "NOT for" line.

Numbered, runnable steps.

The mistakes, why they happen, the fix. #+end_src

• See it discovered: Invoking a toolkit