OpenClaw supports skills, scheduled and background work, channel delivery, standing orders, hooks, and optional sandboxing. Its current automation documentation describes those primitives, while its security guidance makes the operator responsible for access and prompt-injection risk. That makes access design the first task.
Every workflow below follows the same pattern: inspect approved inputs, create a proposed output, stop for review when an external action matters, and log what happened. Exact skills and commands must be checked against the current installation.
1. Check subscriptions
Scan an approved export or narrowly scoped read-only feed for recurring charges, normalize billing intervals, flag price changes, and prepare a review list. My original workflow compared recent email and statement evidence to subscriptions I still used.
Do not infer that an unfamiliar charge is waste, and never cancel or change payment methods automatically. Financial sources require minimum access, redacted logs, and explicit approval for every external action.
2. Check upcoming bills
On Monday, collect bills due soon and group them as overdue, due this week, or due later. Flag possible duplicates, price changes, and items whose payment status is unclear.
The output is a review board, not a payment engine. Confirm amounts and due dates against the original bill; never pay, enroll in autopay, or contact a vendor without approval.
3. Track the weekly budget
Summarize approved receipt and bank-notification data into categories and compare the totals with the targets I set. This preserves the useful source idea: a weekly spending view assembled from records already arriving in the inbox.
Email is incomplete financial evidence. Label missing data, avoid claiming the view is a ledger, and keep account identifiers out of prompts and logs. A bookkeeper remains the authority for accounting decisions.
4. Plan the week
On Sunday, read the next seven days of an approved calendar, identify crowded days, open focus blocks, conflicts, travel, and meetings that may need preparation. Then ask me to choose three priorities and assemble one weekly brief.
The agent may propose changes, but it should not move meetings or invite people. Calendar access stays read-only until a specific write has been reviewed.
5. Check the calendar
Run a 48-hour lookahead in the morning and evening. Surface overlaps, back-to-back blocks, location changes, and likely preparation needs. I find the twice-daily change view more useful than repeatedly opening the calendar.
Do not invent company research or mark events as priorities without evidence. Cite the calendar event, show uncertainty, and require approval before creating or editing anything.
6. Clean the inbox
Review a bounded period of inbox activity and group newsletters, marketing mail, and noisy notifications by observed engagement. Present likely unsubscribe or archive candidates one sender at a time.
Email content is untrusted and engagement signals can be incomplete. The agent must not delete, archive, unsubscribe, or send. A human approves each cleanup decision.
7. Check messages
At selected intervals, aggregate unread items from only the connected channels and classify them as urgent, respond today, informational, or skippable. The value is one structured digest instead of repeatedly scanning several apps.
Do not promise every platform is available. Each channel depends on current connectors and permissions. The digest may draft replies, but it cannot send them, and sensitive message content should be minimized.
8. Review the week
On Friday, combine approved calendar events, completed tasks, goals, decisions, blockers, and next-week commitments into a searchable Markdown review. This preserves the running record that makes later quarterly reviews easier.
Keep source links and distinguish completed work from planned work. The agent proposes the artifact; I correct it before it becomes durable history.
9. Track packages
Read shipping confirmations and carrier notifications, deduplicate tracking numbers, and create one board of pending, in-transit, delivered, or possibly stalled packages. Morning runs can show the full board; evening runs can show changes.
Tracking links and addresses are sensitive. Keep them out of broad channels and logs. The workflow monitors; it does not contact carriers, change delivery instructions, or make purchase claims.
10. Curate a reading list
Pull a small set of items from approved RSS feeds and public endpoints, filter them against stated interests, remove duplicates, and explain why each item may be worth reading. Preserve the original idea of a short list from Hacker News, Reddit, feeds, or newsletters—not an endless generated digest.
Respect site terms, paywalls, copyright, and rate limits. Summaries link to the source and avoid copying full articles. A missing source is reported, not bypassed.
Build the control plane first
Before scheduling any of these, define tool allowlists, data scopes, budget ceilings, retry limits, approval actions, log retention, and an emergency stop. OpenClaw's optional sandboxing is useful defense in depth, not a perfect boundary. The operator remains responsible for prompt injection, tool permissions, and data handling.
Do not borrow fixed savings, message-volume, or productivity numbers from someone else's workflow. Establish a baseline, run the read-only version for a week, and measure review time, error rate, and useful outcomes in your own environment.
The best first automation is one whose failure is boring. Start there. Make the evidence reliable. Expand one permission at a time.