schedule-setup

name	schedule-setup
description	Walk the user through creating the packaged Cowork scheduled tasks (morning briefing, inbox sweep, meeting-prep watcher, follow-up watcher, weekly review). Activate on "/assistant:schedules", "set up my scheduled tasks", "automate my briefing", or "run this every morning".

Schedule setup

Turn the plugin's recurring jobs into scheduled tasks. The canonical job list lives in scheduled/schedules.yaml — five jobs, stable job_id slugs, shared prompts across local Cowork schedules, Claude Code cloud schedules, and managed-agent cookbooks.

Walk the user through the decision tree, then help them create each job they want.

How to create a scheduled task in Cowork

Type /schedule (or use the Scheduled sidebar), paste the prompt, pick the cadence. Each run is a full Cowork session with access to files, connectors, plugins, skills, and the web.

Platform note: local surface requires Claude Cowork desktop (/schedule). Cursor plugin users without Cowork should use cloud-code (Claude Code at claude.ai/code/scheduled) or managed cookbooks for always-on jobs.

Read this to the user before they start

Three honest caveats — say them up front:

1. The machine must be awake. Scheduled tasks on the local surface run on the desktop: Claude Desktop must be open and the computer not asleep. Missed runs fire when the machine wakes — there's no cloud catch-up. If a job is mission-critical (the 8am brief must land), consider cloud-code or the managed-agent version in managed-agents/ which runs on Anthropic's infra.
2. Pre-approve the tools. So a run doesn't stall on a permission wall while unattended, approve the tools it needs ahead of time (e.g. Gmail draft creation, calendar read, file write).
3. Each run uses quota. Batch related work into one task rather than many small ones. After the first run, Cowork may rewrite the prompt from what it learned — expected.

Decision tree

Read scheduled/schedules.yaml and walk the user through surface selection per job. Default is always local — managed agents are opt-in when reliability outweighs local-first.

1. Ask: "Does any job need to run while your laptop sleeps?"
2. If no — recommend local (Cowork scheduled tasks) for all chosen jobs. State the machine-awake caveat once.
3. If yes — for each job they want always-on:
   • reliability_tier: critical (morning-briefing): explicitly recommend managed (managed-agents/morning-briefing/agent.yaml) or cloud-code when sleep is common. Do not imply the whole setup must move — only this job.
   • Jobs with a managed cookbook (inbox-triage-am, inbox-sweep, follow-up-watcher): offer managed or cloud-code as the reliability upgrade.
   • Jobs without a managed cookbook (meeting-prep-watcher, weekly-review): recommend cloud-code (Claude Code scheduled tasks at claude.ai/code/scheduled) — same packaged prompt, Anthropic infra, no new cookbook required.
4. When recommending managed, include the cost caveat from managed-agents/README.md (~$0.08/session-hour runtime; burst schedules are cheap; 24/7 polling is not).
5. Never imply My Assistant hosts schedules — the user deploys Cowork, Claude Code, or managed-agent cookbooks themselves.

Record the user's chosen surface per job_id — you will write it to the health ledger below.

Full user-facing copy of this tree: docs/guide/07-always-on-reliability.md.

Initialize schedule health ledger

After the user picks jobs and surfaces, create or update one file per job at {working-folder}/scheduled/{job_id}.yaml:

version: "0.1"
job_id: <job_id>
updated_at: <ISO-8601 now>
surface: local | cloud-code | managed   # user's choice
cadence: "<cron from catalog>"
last_run_at: null
last_run_status: missed
expected_artifact: null                 # or resolved path pattern for dated artefacts
artifact_present: false
miss_count_7d: 0
notes: null

• Create scheduled/ on first write in the working folder only — never under the plugin directory.
• Write one YAML file per configured job; filename must match job_id (e.g. morning-briefing.yaml).
• Include only jobs the user actually set up.
• If a job file already exists, merge updates; do not wipe existing heartbeat history.

The packaged tasks

Prompts below match scheduled/schedules.yaml. Anchor headings are stable packaged_prompt_ref targets.

Morning briefing {#morning-briefing}

Weekdays 8:00am · job_id: morning-briefing · reliability_tier: critical

Each weekday at 8am, give me a short, scannable briefing: today's
calendar with attendees, important unread emails or messages, follow-ups
going cold, and anything needing my attention before noon. Read my
profile per rules/paths.md first. Run
/assistant:brief. Keep it under 400 words. Save it as brief-YYYY-MM-DD.md.

Cron: 0 8 * * 1-5

Inbox triage (morning) {#inbox-triage-am}

Weekdays 8:00am · job_id: inbox-triage-am

Full triage of new mail since the last sweep: bucket into needs-reply / FYI /
marketing / VIP using /assistant:inbox triage and my profile. Draft replies
for needs-reply + VIP via email-drafting (Gmail drafts when connected).
Extract non-reply actions per actions.policy.md (calendar holds, tasks).
Summarise long threads. Use batch digest tables when FYI or marketing count >= 3.
Run Pass B lingerers scan (read mail older than 3 days with action heuristics).
Save the report as sweep-YYYY-MM-DD-0800.md in my working folder.
Run: python3 scripts/update_ledger.py --job-id inbox-triage-am --status success
  --artifact sweep-YYYY-MM-DD-0800.md --working-folder {working-folder}
Before ending, confirm last_run_at in scheduled/inbox-triage-am.yaml matches
this run. List what you propose archiving.

Cron: 0 8 * * 1-5

Inbox sweep (midday/afternoon) {#inbox-sweep}

Weekdays 12:00pm, 4:00pm · job_id: inbox-sweep

Light sweep of new mail since the last run into needs-reply / FYI / marketing /
VIP using /assistant:inbox sweep and my profile. Summarise long threads.
Use batch digest tables when FYI or marketing count >= 3. Skip reply drafting
except Tier 1 VIP needing immediate attention. Skip Pass B lingerers unless asked.
Save the report as sweep-YYYY-MM-DD-{slot}.md in my working folder
(slot 1200 or 1600 by local time). Run: python3 scripts/update_ledger.py
  --job-id inbox-sweep --status success --artifact sweep-YYYY-MM-DD-{slot}.md
  --working-folder {working-folder}
Before ending, confirm last_run_at in scheduled/inbox-sweep.yaml matches this run.
List what you propose archiving.

Cron: 0 12,16 * * 1-5

Meeting-prep watcher {#meeting-prep-watcher}

Weekdays 7:00am · job_id: meeting-prep-watcher · no managed cookbook — cloud-code fallback

For each meeting today with an external attendee, run /assistant:meeting prep:
who they are, what they do, our last contact, and what I should prepare.
Read my profile and memory first.

Cron: 0 7 * * 1-5

Follow-up watcher {#follow-up-watcher}

Daily 5:00pm · job_id: follow-up-watcher

Check mail I sent that's awaiting a reply using /assistant:email review.
For anything silent 3+ business days, draft a polite nudge (do not send)
in my voice and list it. Cross-check the Blocked section of TASKS.md.

Cron: 0 17 * * *

Weekly review {#weekly-review}

Friday 4:00pm · job_id: weekly-review · no managed cookbook — cloud-code fallback

Run /assistant:report: summarise the week — what closed, open loops,
stale tasks (untouched 14+ days), and what to line up for Monday. Update
TASKS.md and prune stale memory with my confirmation. Save as
review-YYYY-MM-DD.md. Keep it scannable.

Cron: 0 16 * * 5

Suggested starter set

If the user wants a minimal setup, recommend just two: Morning briefing and Inbox sweep. Add the others once those prove useful.

Health recording on scheduled runs

When a scheduled run completes (any surface), the target skill updates scheduled/{job_id}.yaml for its job via scripts/update_ledger.py — do not hand-edit the ledger:

python3 scripts/update_ledger.py \
  --job-id {job_id} \
  --status {success|partial|failed} \
  --artifact {artefact-path-if-any} \
  --working-folder {working-folder} \
  [--notes "optional context"]

The script sets:

• last_run_at to now (ISO-8601)
• last_run_status to success only when required artefacts exist or the skill's completion criteria are met; otherwise partial or failed
• expected_artifact and artifact_present when an artefact path is supplied
• updated_at on the job file

Preserves existing surface, cadence, and miss_count_7d. Interactive runs do not reset miss_count_7d unless they complete the scheduled job's artefact contract.

Handoff

Point the user to docs/guide/07-always-on-reliability.md for the full reliability story, escalation triggers, and cost caveats.

Approval frame

Follow rules/approval-frame.md when a scheduled run surfaces reviewable work.

Queue mappings:

• Cron configuration alone does not create queue items.
• Inbox sweep runs that propose archives write archive-proposal items (one batched item per sweep) per inbox-triage conventions.

Append the observability footer when queue items are written.