Scheduled briefs (routines)

Run Chief of Staff on a schedule so briefs arrive automatically instead of only on request. The mechanism depends on the host (see brain/integrations/hosts.md):

• Claude Code desktop — Routines. Chief of Staff creates these itself using its routine tools (the principal approves each tool execution) — see "Setting them up".
• Codex (in the ChatGPT desktop app) — Scheduled tasks (the Scheduled tab). Set one per brief with the same cadence
• prompt (below); guide the principal to add them, or add them if the host lets you.

Either way the briefs **read memory/, so they must run locally against this folder** — not a cloud/remote copy (which wouldn't have this principal's memory/; see "Setting them up"). The cadences below are defaults the principal can override or turn off (see the Briefings section of brain/role/defaults.md).

Recommended routine set (defaults)

Briefs — regular deliverables:

Maintenance — keep memory clean and connected (read-only reports that propose changes, they don't edit memory on their own):

Each routine runs in the Chief of Staff project, so it loads the brain, reads memory/, uses the connected calendar, and produces the brief. The maintenance routines also append a run entry to memory/ledger.md (per brain/schemas/capture-rules.md → "Action ledger"), which is what the session-start maintenance offer reads to know when they last ran (see CHIEFOFSTAFF.md boot protocol).

Setting them up (onboarding or later)

Run this flow at onboarding step 6, on /routines, or whenever a brief would help. It applies to any scheduled task Chief of Staff creates — the five default routines, per-member routines, and any ad-hoc one-off (e.g. "remind me every Friday to send the update").

1. Review what's already scheduled first — don't create blind. Before proposing or creating anything, take stock of the current scheduled tasks: read the registry in memory/preferences/briefings.md (step 6 keeps it current — it's the reliable source), and if the host exposes a way to list existing routines/scheduled tasks, use that too. Hold that picture for the dedup check in step 3.
2. Propose all five at once as a compact menu with their default cadences (the tables above) — the three briefs plus the two maintenance routines (memory lint monthly, explore biweekly, each with its one-line benefit) — and ask for yes / adjust the times / skip. Frame them as good defaults.
3. Before creating each task, check it against what's already scheduled — suggest a dedup on overlap. Overlap = the same purpose (same playbook/brief) or an overlapping cadence/time. If a task would duplicate a current one, don't silently add a second — surface it and offer a choice: update the existing one, adjust its timing, merge them, or skip. Let the principal decide. If you can't tell for sure whether it already exists (empty registry and no way to list the host's tasks), ask once rather than assume ("looks like you may already have a daily brief — want me to update that one instead of adding a second?").
4. Create each one they keep — how depends on the host (brain/integrations/hosts.md):
• Claude Code: create it yourself with your routine tool — don't make them open the UI. Create a persistent Local routine (runs in this project → reads memory/, brain, calendar) with the schedule + prompt = run the matching playbook. The principal approves the tool execution — that's the confirmation. Use a persistent (not session-only) routine.
• Codex: add each as a Scheduled task (Scheduled tab) with the same cadence + prompt — do it if the host lets you, otherwise hand the principal the exact cadence + prompt to add. It must run locally against this folder so it can read memory/.
• Other host with no routine tool: walk the principal through its scheduler (or a headless run + system cron), Name / Schedule / prompt = the routine prompt / working folder = this project, running locally.
5. Changing or removing one later (different time, turn one off) — update/delete it the same way for the host (your routine tool on Claude Code; the Scheduled tab on Codex). When the principal asks for a change, prefer updating the existing task over creating a new one.
6. Record every scheduled task you create in memory/preferences/briefings.md — its cadence, the prompt/purpose, the host, and status (created, declined, or deferred) — plus any cadence change as an override (per brain/role/defaults.md). Recording every task, not just the five defaults (ad-hoc ones too), keeps this file a complete, enumerable registry so step 1's review can dedup reliably even on a host that can't list its own tasks. This is also what lets Chief of Staff stop re-asking (see CHIEFOFSTAFF.md → CONNECTORS & ROUTINES) and what the session-start maintenance offer checks.

• Why local (not cloud/remote): a cloud/remote run happens on a fresh copy with NO access to this principal's memory/ — so any memory-touching brief would see empty memory and produce an empty brief. A local run (Claude Code Routine, or a Codex Scheduled task on this folder) reads memory/, the brain, and the connected calendar. This is true on every host.
• Let them change the times, or disable any — these are defaults, not requirements.

Per-member routines (virtual team cadence)

If the principal runs a virtual team (brain/schemas/team-member.md), a member's cadence: duties become Local routines too, set up the same way (propose → create with the host's routine tools / a Scheduled task, gated by the tool-approval prompt). A member routine's prompt is: "Act as <slug> (load memory/team/<slug>/charter.md + its memory + the shared brain in code names), run this standing duty, then hand the result to Chief of Staff to verify and integrate." It obeys the member's autonomy (draft-and-wait ⇒ prepare and surface only — nothing outward/irreversible), must run locally (it reads the member's own memory), and records its run in the member's ledger.md. On a run is a persona-swap, the same on every host.

Honest limits

• Scheduled briefs are what make auto-delivery possible — without them, briefs only run when asked.
• On Claude Code, you create routines yourself with your routine tools (not silently): each create/update surfaces a tool-approval prompt the principal accepts (and a one-time working-folder trust prompt may appear). On Codex, briefs are Scheduled tasks.
• A scheduled brief typically only fires while the app is open and the machine is awake (host behavior varies — e.g. Claude Code runs a single catch-up for a recent missed time). Each run arrives as a notification / new session — it isn't emailed.
• It must run locally against this folder — a cloud/remote run can't see the principal's memory/, so it would produce an empty brief. Verify a new host's scheduler runs locally (brain/integrations/hosts.md).
• A brief with no calendar connected still runs; it just asks for / omits the schedule (see brain/integrations/calendar.md).