The general agent-guidance skills — collab, build, resource-create,
resource-update, story, and watch-requests — and the system that
composes them.
Each skill is a template. The renderer splices in shared doctrine from
fragments/ and the operator's own voice from ~/.config/agentguidance/, then
writes the read-only artifacts every configured agent loads.
These skills are the general layer of the agent* fleet: how to read a
request, when to sketch and wait, how to build well, how research becomes a
durable resource, how a codebase becomes a story. Tool-specific runbooks live
with their tools — agentboard ships board, agentwiki ships wiki. This
repo carries only the doctrine that applies everywhere.
AgentStart owns installation. Its skills scan ships skills/<name>/ whole, then
runs scripts/post-sync — the render — so the installed copies arrive already
composed with the operator's extension prompts. Nothing here is installed by
hand. A machine without the AgentStart seam can still render directly:
scripts/render # templates + fragments + ~/.config/agentguidance → ~/.agents/skillsskills/<name>/— one template per skill:SKILL.md(with fragment and extension-prompt render points),agents/openai.yaml, and any reference files, shipped whole.fragments/— doctrine shared between templates; a missing fragment fails the render.scripts/render— the composer;scripts/post-sync— the fleet hook that execs it.tests/validate.sh— the gate; run it before committing.
The extension prompts themselves are not here: they are the operator's,
kept in AgentStart (prompts/agentguidance/) and linked into
~/.config/agentguidance/. That split is the point — these templates are
public doctrine, the extensions are one machine's voice, and the render is
where they meet.