agent-brief-hygiene

Agent brief hygiene

What actually reaches a dispatched subagent

A plain dispatched subagent (not a "fork" of the current session, and not one of the built-in Explore/Plan agent types) starts with the full CLAUDE.md hierarchy the main conversation itself loads — user-level, project-level, and any directory-level CLAUDE.md files — as part of its initial context, automatically, with no action required to hand it over. A rule that lives directly in the installed CLAUDE.md (the Role boundary, Model routing, Guardrails, and so on) is already in front of that subagent before its brief is even read.

What genuinely does not transfer to a dispatched subagent: the coordinator's conversation history (prior messages, what other agents already found), a skill's body that the coordinator itself invoked (loading a skill is scoped to the session that invoked it — a subagent doesn't get it merely because the coordinator does, and won't spontaneously invoke a sibling skill it doesn't know exists), auto-memory, output style, and the contents of files the coordinator has already read but the subagent hasn't.

Two dispatch shapes break the general rule above, in opposite directions: a fork clones the parent session's own context, conversation history included, so it needs none of this restated; the built-in Explore and Plan agent types skip CLAUDE.md entirely, so treat a dispatch to either of them like the plugin-only case below for every rule that would otherwise arrive for free.

What still needs restating, and why

Two different reasons produce the same-looking instruction — "restate this in the brief" — and it matters which one applies, because only one of them is actually about inheritance:

• A rule that lives only in a plugin skill is not inherited at all. coordinator-kit's credential-handling, backlog-discipline, comms-register, and question-protocol skills (and this one) are not part of the CLAUDE.md hierarchy — a subagent gets none of their content unless the brief either pastes the operative rule inline or explicitly names the skill for the subagent to invoke itself. The same is true of execute-loop's push gate: it is task-specific dispatch-loop mechanics, not a standing CLAUDE.md rule, so it reaches a build agent only if the brief pastes it in. For anything safety-critical (credential handling above all), do both: name the skill and paste the specific operative rule inline rather than trusting that a subagent will think to invoke it.
• A rule that lives in the installed CLAUDE.md already reached the subagent — restating it is about salience, not delivery. CLAUDE.md is long; a generic slot like Guardrails is filled in with this project's actual specifics (which URL is production, which action is irreversible) and a build agent shouldn't have to search a few hundred lines to find the one paragraph that applies to its task. Naming the concrete, already-filled-in fact directly in the brief is what makes the standing rule actually operative for this one task, not what makes it reach the subagent in the first place.