Four recurring AI jobs people actually run, with prompts to copy

Pick one job you repeat every morning, every week, or every release path and copy one of these. Each build below is someone's own recurring run, cited to their posts, with the task, the tool, the result they report, and a starter prompt or request body. Results are self-reported. We have not reproduced these runs, and every prompt marked "ours" is ours and untested. Where a builder sells something related, we say so in their block.

1. A CFO's five Claude Code scheduled tasks (WhatsApp to ledger)#

Biel Rosinol (@gabrielrosinol). His bio says "CFO & fintech founder. Rebuilding the finance function for AI, in the open." He is in Barcelona. Disclosure: his site, bielr.com, has a "Work with me" link to contact, so he has a consulting interest in showing this off. The thread itself sells nothing.

• Task: the recurring life and finance admin a busy person forgets: replying to messages, renewal deadlines, checking the books, watching Claude usage, filing downloads.
• Tool: Claude Code scheduled tasks. On Oct 24, 10:26 AM ET he wrote: "I've been running a homemade version on Claude Code scheduled tasks. 5 routines, no new app." He frames it as his own check-with-me-first version of Claude Duet.
• The five (times as he wrote them):
  1. WhatsApp sweep, 08:15 and 20:15: "Replies and payments come back prepared, and I tap to send. Nothing goes out on its own."
  2. Deadlines sync, day 1 of the month: insurance, ITVs (Spain's vehicle inspection), IBI (Spain's property tax), leases and IDs in one sheet; "The routine rebuilds my calendar and task list from it."
  3. Ledger review, day 1 of the month: double-entry personal finances since 2017; "The routine checks every control in the ledger and tells me what doesn't tie."
  4. Plan pace, 23:00 every night: "Logs my Claude usage against an even pace to the weekly reset."
  5. Filing, Sunday 10:00: "Everything I downloaded during the week goes into the right Drive folder."
• Result (his words): On the deadlines post he wrote: "The calendar became an output. I stopped typing reminders." Asked which earns its keep, he replied Oct 25, 9:56 AM ET that the WhatsApp sweep is "the one I'd notice first if it stopped" and "The ledger review is the one I'd least like to go back to doing by hand."

How to copy it

Claude Code has two kinds of scheduled tasks (Desktop docs):

1. Local task (Claude Desktop): "runs on your machine with direct access to your files and tools, but only fires while the app is open and your computer is awake." Use this for jobs that read Downloads or other local files, like his filing job.
2. Cloud task (formerly routines): "runs even when your computer is off," and can also start on API calls or GitHub events (cloud docs). Connector actions "appear as you."
3. In Desktop, open Scheduled tasks, create a task, pick Local or Cloud, paste the prompt, set the schedule, and choose the permission mode.
4. Click Run now once and approve each tool so later runs do not stall. Docs: "To avoid stalls, click Run now after creating a task, watch for permission prompts, and select 'always allow' for each one."
5. Keep the sending step manual. Do not grant a send tool for anything that messages a person; have it draft. That step is our addition to his rule, turned into setup.

Our starter prompt (ours, untested; modeled on his deadlines sync; adjust the sheet name and calendar):

Open the sheet "Deadlines" (columns: item, type, due date, notes).
For every row due in the next 90 days:
1. Make sure there is one calendar event on the due date and one
   reminder 14 days before it. Create what's missing; don't duplicate.
2. Add a task to my task list for anything due in the next 30 days.
3. If a due date in the sheet changed since last month, move the
   matching event instead of creating a new one.
Do not email, message or pay anyone. At the end, list what you
created, moved or couldn't match, in under 10 lines.

Gotchas

• A local task that sleeps through its time is skipped; on wake, Desktop starts "exactly one catch-up run for the most recently missed time" and discards older misses. An 08:15 sweep can fire much later; put a time check in the prompt (Desktop docs).
• Scheduled tasks "draw down subscription usage the same way interactive sessions do," which is why his nightly plan-pace check exists (cloud docs).
• Cloud tasks are a research preview: "Behavior, limits, and the API surface may change."

Caveats: self-reported; no time-saved numbers; 83 followers as of Oct 28. Credibility rests on the specificity of the thread, his public bio, and the consulting disclosure above. Distinct from the Oct 23 grill-me workflow (different author, tool, and task).

2. A morning cleanup of bot followers (Grok Bot)#

Ashley Schendel (@ashleyschendel). Her bio says "Tennessee mom building Grok Bots for real life. Sports, school, grocery bills and what's for dinner." About 8,030 followers as of Oct 28. Disclosure: she shares Grok Bot templates (her X link opens a bot.ski templates tab), and Bot Terminator is a free marketplace template by Empress Heavy (@HeavyMetalShip), "Grok Bot Dev of Bot Terminator + DOGE - Available on Grok Bot Marketplace." Her posts describe her own use. They do not sell the template.

• Task: clear bot and spam followers off X without spending the morning on it by hand.
• Tool: Grok Bot running the free Bot Terminator template on a daily routine. The template page: cleans "bot, spam and long-inactive followers," can "remove your account from unwanted lists on X," and "You review and decide what happens."
• Result (her words): On Oct 26, 3:26 PM ET she wrote: "I've been running @HeavyMetalShip's Bot Terminator every morning, and it removes 10 bot followers a day for me." On Oct 28, 9:42 AM ET: "I already use Bot Terminator by @HeavyMetalShip every morning to clean up my followers" (post). Same morning at 10:51 AM ET: "I have it set up on a routine to remove 10 a day" (post).

How to copy it

1. Add the free Bot Terminator template, or create a Bot with the same job.
2. Connect X only. Keep post and message tools off for this Bot.
3. Ask for a weekday morning routine with a small daily cap (she uses 10). Per routines docs, confirm the schedule card, then Test once on desktop. Docs: "A test run performs real work."
4. Require approval on every removal for the first week.

Our starter prompt (ours, untested; modeled on her 10-a-day routine and the template's review rule):

Every weekday morning, scan my X followers for clear bot and spam
accounts. Propose at most 10 removals. For each one, show the handle,
why it looks like a bot or spam, and a link to the profile. Wait for
my yes before removing anyone. Never block, mute, post, or message
anyone. Never touch people I follow or accounts I've replied to in
the last 90 days. When you're done, send me a short list of who you
removed (if I approved any) and who you skipped.

Gotchas

• Template says review first; on Oct 28, 12:57 PM ET she replied: "If it's clearly a bot, I let it remove it automatically." Start with always-ask.
• Removals are hard to undo at scale. Keep the cap small.
• Routine runs spend weekly usage (docs). Do not stack hourly scans.

Caveats: self-reported; no run-log screenshot in the cited posts; 10-a-day is her number. Distinct from email/SMS spam cleanup tips (Sunny Sahota's 5am inbox candidate was held from the Oct 22 showcase; the Oct 22 FoundAt5280 tip is texts/email, not X followers).

3. A one-page prep before the daily team sync (Grok Bot)#

Marko Marinovic (@mmarinovic_). Bio: "building systems that move money | fintech, blockchain, AI | father of two | co-founder & cto." His site, markomarinovic.com, says he is a staff engineer and team lead at Lisk, and that through Blank he "consult[s] for companies building payments and AI systems." Disclosure: consulting interest. The Oct 21 post sells nothing and includes a copy-ready prompt card.

• Task: walk into the daily team sync prepared without reading calendar, Notion, Slack, Linear and email by hand first.
• Tool: a Grok Bot routine every weekday, 30 minutes before the sync. On Oct 21, 4:47 AM ET he wrote: "I don't prep for the daily team sync anymore. My Grok Bot does." It reads "my calendar, today's agenda in Notion, the latest planning notes, Slack since yesterday, Linear progress and comments, cross-team threads and the engineering digest."
• Result (his words): "hands me one page" with time and link, the agenda, "release scope in five lines or less," "one line per teammate," and "three things worth raising." Same post: "If a bot can write the status, nobody should read it out loud. Spend the daily team sync on the things worth raising."

How to copy it

1. Connect calendar, Notion, Slack, Linear and work email to one Bot for this job only (routines docs).
2. Paste his prompt, fill <your daily team sync time>, <Notion agenda database> and <team>.
3. Schedule weekdays, 30 minutes before the sync. Confirm the card. Test once on desktop.
4. Keep Slack posts and Linear writes behind approval. This prompt only reads and writes you a page.

His starter prompt (verbatim from the copy-ready card on his Oct 21 post):

# Prep me for my daily.

Run this every weekday, 30 minutes before <your daily team sync time>.
Stay silent if the daily team sync is cancelled or declined. Send one
line if it's async today.

## 01 / Gather

1. Check the calendar for today's daily team sync. Grab the meeting
   link and any daily team sync plan.
2. Find today's agenda page in <Notion agenda database> and read its
   host and agenda items.
3. Pull release scope and dates from the latest planning notes. Check
   Slack for any scope or date changes leads posted since the last run.
4. For each person on <team>, pull their Linear progress since
   yesterday: issues moved, blocked, or newly assigned.
5. Pull important Linear comments on those issues since yesterday.
6. Pull important Slack messages and threads aimed at <team>, or
   discussing the team's work, since yesterday.
7. Find cross-team Slack threads about features <team> is working on.
   For each, give the latest update in one line.
8. Pull the engineering digest and new work emails since yesterday.

## 02 / Format the result as

- time and meeting link
- agenda page and host
- agenda items
- at most 5 lines of release scope
- each teammate's progress (one line each)
- important Linear comments (at most 5)
- important team threads (at most 5)
- cross-team feature threads (one line each, with the latest update)
- up to 3 prep bullets

## 03 / Rules

Never invent status. If a source is empty or unavailable, say so in one
line and move on.

Gotchas

• Connectors must be signed in to the right workspace. Empty source: say so in one line (his rule).
• Cancelled or declined sync should stay silent. Odd calendar labels need a schedule check.
• Routine runs spend weekly usage. One weekday run is the point.

Caveats: self-reported; 122 impressions on the source post as of Oct 28; consulting disclosure above. Prompt verified from his card image this gate. Distinct from Brad Smith's Oct 22 GHL CRM reports (sales pipeline, not eng standup prep).

4. Live risk scoring for agent shell commands (Decisions API)#

AJ Stuyvenberg (@astuyve). Bio: "Staff eng @Datadoghq" and "(opinions mine)." About 12,850 followers as of Oct 28.

• Task: check how risky an agent-generated shell command is before it runs on a host. Per Datadog's docs, rshell "is the command execution layer behind the remote-actions toolset in the Datadog MCP server."
• Tool: OpenAI's Decisions API (we covered it at GA in our Oct 22 story).
• Result (his words): On Tuesday Oct 27, 12:06 PM ET he wrote: "We're using the Decisions API to help score RShell commands used to remediate issues on hosts and clusters. Previously we did this offline, but it's so fast and accurate enough (with a good confidence score) so we can actually do it live now." A follow-up at 12:31 PM ET attaches the latency table and says the comparison is an "internal service that can use multiple frontier models to score risky code and shell commands," and "it was closer to 40.3x at p95." Table (OCR this gate): p95 20,451 ms Frontier Model vs 507 ms Decisions. OpenAI's launch post already quoted him: "it delivered 40x faster p95 with comparable accuracy." @OpenAIDevs reposted the Oct 27 post at 1:35 PM ET.

How to copy it

Per OpenAI's Decisions guide, send POST /v1/decisions with model, input and questions. A score question rates input "against ordered levels." The guide says to "Choose thresholds based on the cost of false positives and false negatives." Only gpt-6-luna is supported.

Request body (ours, untested), in the shape of the guide's score example:

{"model":"gpt-6-luna","input":"Proposed fix: truncate -s 0 /var/log/app/debug.log","questions":[{"type":"score","name":"risk","instructions":"How risky is it to run this command on a production host?","levels":[{"label":"Safe","description":"Read-only or confined to log files."},{"label":"Review","description":"Changes state but is reversible."},{"label":"Block","description":"Destructive or irreversible."}]}]}

Gotchas

• Use the confidence score as the escape hatch. Reply, 12:23 PM ET: "When it DID hit a false positive or false negative the confidence score was more like 50% (so I could pass it off to an LLM or person for analysis)."
• Test side by side first. Same reply: "I ran it in parallel with Jev and a jev-like model we self hosted. I found Open AI's confidence score to be the most useful." We could not identify "Jev"; leave it quoted as written.
• One model. Guide: only gpt-6-luna is supported.

Caveats: reported / self-reported from a Datadog staff engineer. OpenAI's Oct 22 launch post already quoted him on rshell scoring; the new material is the Oct 27 offline→live post, the full latency table, the parallel test, and the 50% confidence tip. Our GA story quoted Sierra, not Datadog. Distinct from the Oct 22 Dot showcase block.

What to watch#

Three of these four builds keep a human on the send or remove step: Biel drafts WhatsApp replies, Ashley starts with approval on removals, Marko's prompt only reads and writes a page. Copy that first. The Datadog build is the one that runs in a live path, and Stuyvenberg's tip is to treat a ~50% confidence score as a handoff, not a green light. Scheduled and routine jobs also spend subscription usage; Biel's nightly plan-pace check and the Grok Bot usage gotchas are the reminder not to stack hourly scans.

Sources#

• Biel Rosinol, Oct 24, 2026 (10:26 AM ET), WhatsApp sweep, deadlines, ledger, plan pace, filing, lesson, Oct 25 which earns its keep (9:56 AM ET), Desktop scheduled tasks, cloud routines, bielr.com
• Ashley Schendel, Oct 26, 2026 (3:26 PM ET), Oct 28 already use (9:42 AM ET), routine 10 a day (10:51 AM ET), auto when clearly a bot (12:57 PM ET), Bot Terminator template, routines docs
• Marko Marinovic, Oct 21, 2026 (4:47 AM ET), Oct 19 EA team, Oct 20 routine tip, markomarinovic.com
• AJ Stuyvenberg, Oct 27, 2026 (12:06 PM ET), confidence tip (12:23 PM ET), latency table (12:31 PM ET), OpenAIDevs RT (1:35 PM ET), OpenAI Decisions launch, Decisions guide, Datadog rshell