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

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

1. A pizzeria's morning books report (Claude Code)#

Chris Simone (@CSS976) has run a pizzeria since 2003 and wrote on Oct 5: "I've run a pizzeria for 23 years. I don't write code."

• Task: keep the restaurant's books current and know where the money stands each morning, not at month end.
• Tool: Claude Code, which he used to build the system, plus QuickBooks. On Oct 9 he wrote: "I use quick books still to organize it all, but Claude operates the QB, about 9 months in."
• Result: replying on Oct 17 to a CPA who asked how people automated their accounting, he wrote: "Mine’s been automated for about 9 months, built it all in Claude code, runs on a remote server, I get an email every morning, or a text if urgent." His Oct 9 post also claims "23 years of books cleaned up in minutes, balanced to the penny." Those are his claims; we haven't seen the system.

How to copy it

1. In Claude Code, describe the report you want and where the data lives (accounting software, bank export, supplier invoices).
2. Have it build a script that pulls the data, checks it, writes a short summary, and has a rule for what counts as urgent.
3. Run it on a schedule on a machine that stays on, and have it email you the summary and text you only for urgent items.
4. Keep a human approval step on anything that changes the books or moves money. That step is our addition, not his described setup.

Our starter prompt (ours, untested):

Build me a small program that runs every morning at 6 AM. It should pull
yesterday's sales, expenses and bank activity from my accounting software,
check that they reconcile, and email me a one-page summary: sales versus the
same weekday over the last six weeks, food cost, labor, and anything that
doesn't balance. If something looks urgent, like a missing deposit or a
duplicate charge, text me instead. Don't change or post anything in the books
without asking me first, and show me how to run it on a schedule.

Gotcha:

• A retired model can break a build silently. He wrote that an inventory system he built in the spring "died and nobody noticed," because "One line of code pointed at an AI model that had been retired. It sat there broken for two months." He didn't name the model or provider. Our read: have a scheduled build alert you on API errors, and check Anthropic's model deprecations page for retirement dates.

Disclosure: he sells an iPhone app for restaurant numbers, Keep the Dough, which he announced on Oct 14. That thread doesn't say it was built with Claude, and we don't link the two.

2. Morning and evening CRM reports (Grok Bot)#

Brad Smith (@bradsmith), whose X bio reads "Founder of @AutomationLinks," wrote on Oct 16: "For the last 3 years, I've been logging into GHL every day." GHL is GoHighLevel, a CRM used by agencies. "Now my Grok Bot tells me every morning and every night."

• Task: the daily login to check new leads, contacts and appointments.
• Tool: a Grok Bot connected to GoHighLevel. In a reply he said he picked Grok Bot "since it plugs right into GHL and sends me the report every morning and night."
• Result: "It found over 400 potential leads that we could follow up with to remind them about us." That figure is his.
• Repeated use: on Oct 21 he wrote that the Bot is "Managing our entire CRM now," and that evening he replied: "Haven’t logged in over 2 weeks."

How to copy it

1. Connect your CRM to one Bot and give it only this job.
2. Ask for two short reports a day: new leads and today's calls in the morning, pipeline changes and stale leads at night.
3. Keep lead messages as drafts. His post lists "Messages my leads," but asked whether it messages leads on its own, he replied: "Drafts for now but you can always prompt it."

Our starter prompt (ours, untested):

Every weekday at 8 AM and 6 PM, check my CRM. Morning: new leads since last
night, today's scheduled calls, and anyone I promised to follow up with today.
Evening: deals that moved stages, leads with no contact in 14 days, and a
draft follow-up for each stale lead. Save follow-ups as drafts only. Never
send messages, change deal stages or delete contacts without asking me.

Gotcha: his public template is not the report bot. The template page, "GHL Builder by Brad," says it helps you "Set up a GoHighLevel sub-account from a plain-English chat," and that it "was created by a third-party user, not by SpaceXSI." He told one user, "You need my full trained one" (reply).

Disclosure: this is also a promotion. Smith runs an automation agency, his post ends "Leave a comment if you'd like access," and he says he'll be "posting free bots you can use every week."

3. A team's daily CI fixes, moved from personal routines to one workflow (Claude Code)#

Akos (@akoskm), whose X bio says he is a principal engineer at ABBYY, described a pattern "we started doing lately in the team." He doesn't name the team or project, so we don't attribute it to ABBYY.

• Task: fix failing CI builds that "are easily fixable with agents."
• Tool: first, Claude Code routines on each person's Claude desktop app; then Claude Code in GitHub Actions.
• Result: "we create a routine that runs every day with a prompt solving the problem." When they tried to scale it to PR reviews, two things broke: "these routines are local on everyone's claude desktop," so "the reviews or resolutions aren't consistent," and several people's routines could review the same PR at once. Moving to GitHub Actions "solved both: 1. prompt tracking - workflows are committed to the project 2. race condition - a single workflow runs for example on every new commit." He gives no counts or time saved.

How to copy it

1. In Claude Code, run /install-github-app. Per the GitHub Actions docs, it "installs the GitHub App, adds your authentication secret, and prepares the workflow pull request for you."
2. Add a scheduled workflow with a prompt input. The docs warn that "For a plain-text prompt, Claude has no shell or GitHub API access until you grant the tools the prompt needs," with --allowedTools in claude_args, and that "GitHub runs scheduled workflows only from the default branch."
3. Put project rules in CLAUDE.md so every run uses the same standards.
4. Trigger on one event (a schedule or a push) instead of each person running their own routine.

Our starter workflow (ours, untested; adapted from the docs' "Run on a schedule" example; trim the allowed tools to what your repo needs):

name: Daily CI fix
on:
  schedule:
    - cron: "0 13 * * 1-5"
jobs:
  fix:
    runs-on: ubuntu-latest
    permissions:
      contents: write
      pull-requests: write
      actions: read
      id-token: write
    steps:
      - uses: actions/checkout@v6
      - uses: anthropics/claude-code-action@v1
        env:
          GH_TOKEN: ${{ github.token }}
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          prompt: |
            Look at the most recent failed CI runs on the default branch.
            For each failure with a clear, small fix (lint, a broken import,
            an outdated snapshot), open one pull request with the fix and a
            short note on the cause. Skip anything that needs a design decision
            and list it in the PR description instead. Never push to the default branch.
          claude_args: |
            --allowedTools "Read,Edit,Bash(gh run list:*),Bash(gh run view:*),Bash(git checkout:*),Bash(git add:*),Bash(git commit:*),Bash(git push:*),Bash(gh pr create:*)"

Gotchas:

• Bots can't start it by default. The docs say scheduled runs are attributed to "the one who last changed the workflow's cron schedule," and the action rejects a bot actor "unless you list it in allowed_bots."
• Cloud scheduled tasks are another fix. His complaint is about routines on each person's desktop. Anthropic's docs now call them "scheduled tasks, formerly routines," which can "run in the cloud on a schedule, on API calls, or on GitHub events," in research preview.

4. An hourly usage-budget panel, plus a daily brief (Hark)#

LarMedAI (@larmedai) described the setup in an Oct 18 post as being "for me as a P.E. teacher staying up to date with research." The account's X bio says it runs the Lär Med AI YouTube channel, which teaches ChatGPT. We found no tie to Hark.

• Task: stop running out of Hark usage before the month resets, and get a daily rundown without asking.
• Tool: Hark Pro, with a Panel, which Hark's Using Hark page says you can redesign by tapping Reply.
• Result: on Oct 18 the user wrote: "set up a panel like this one. It updates every hour and calculates my daily usage so I can use Hark for the whole month. Also on pro3 after going hard on free to pro2 and now found a good balance on pro3". On Oct 8 the user wrote: "I have it tracking all my subscriptions, insurances, let me know when it's time to renew, compare, present me options without me thinking about it, give me a daily brief where it checks my calendar, mail, potential purchases and more". The panel's numbers are in a screenshot we didn't read, so we quote no figures. These are the user's claims; we haven't seen the setup.

How to copy it

1. Ask Hark to build a Panel that shows your usage against the monthly reset date.
2. Have it refresh every hour and work out a daily budget from what's left.
3. Put it where you'll see it. Another user, TheDesertDog, suggests: "Make a panel, put it on your Home Screen".
4. Before a big task, ask Hark how much usage it expects to take.

Our starter prompt (ours, untested):

Build me a Panel that updates every hour. Show how much of my monthly Hark usage I've used, how many days are left until it resets, and how much I can use per day from here to last the month. Turn it red if today's use is ahead of that pace.

Gotchas:

• The daily bar isn't a daily allowance. Brett Adcock, Hark's founder, says one day can use up to a fifth of your monthly usage; see our pacing explainer.
• Plans changed along the way. The user wrote on Oct 13 about going "from Pro1 to Pro2 and I'm now on Pro3", so the panel's budget depends on your plan.

5. A consultant's dot as his executive assistant (ChatGPT dots)#

Treb Gatte (@tgatte), whose X bio reads "CEO @marqueeinsights" and "Microsoft Data & AI MVP," uses a ChatGPT dot, OpenAI's always-on agent, as his assistant. Disclosure: he sells consulting on this choice. His post ends: "If you're weighing whether to build your own or start with a dot, send me a message and we'll work out which fits your team."

• Task: keep up with meeting changes, coordinate the day and catch urgent client mail.
• Tool: a ChatGPT dot with his email and calendar connected.
• Result: on Oct 5 he wrote: "My @OpenAI dot flagged a client's high-priority email and summarized it before I asked." He then priced out a comparable assistant built on Microsoft and Anthropic and concluded: "For a personal assistant, I'm keeping my dot." His cost figures for those builds are, in his words, "modeled costs, not measured usage."
• Repeated use: "I've had my dot running since #DevDay as my executive assistant. It keeps me updated when meetings change and helps coordinate the day." A day later he replied: "It’s my executive assistant," adding that it took "the three minutes it took to set up my Dot." OpenAI announced dots on Sep 29, so that was about a week of use when he posted.

How to copy it

1. Check you're eligible. Per OpenAI's help article, dots are "rolling out" to Pro users 18 and older outside the European Economic Area, Switzerland and the UK, and are available to Business Premium users; Enterprise admins must turn on the beta.
2. Create your dot in the ChatGPT mobile app, desktop app or desktop web, and "Follow the onboarding prompts to get started."
3. Connect your email and calendar. The article says you "can connect your personal email account" and can ask the dot to "review your calendar each morning."
4. Set a custom rule so it asks before it sends anything. The article lists "Ask before taking action" as one of the behaviors you can choose.

Our starter prompt (ours, untested):

Be my executive assistant on weekdays. Each morning at 7 AM Eastern, read my
calendar and inbox and send me a short plan for the day. Tell me right away
when a meeting moves or a client marks something urgent, with a two-line
summary. Draft replies for my review, but never send email or accept invites
without asking me first.

Gotchas:

• Everything lands in one thread. On Oct 5 he wrote: "Everything is in one massive stream which makes for endless scrolling looking for something we discussed." OpenAI's help article now says "Each task has its own conversation," so check whether that fixes it for you.
• Disconnecting doesn't erase. The help article says "Disconnecting an app does not delete information your dot has already obtained from it." To delete it, you have to reset the dot.
• Price is his figure. He gives Business Premium as "$100 per user per month billed annually, or $125 monthly." OpenAI's article says only that "Your first dot is included in your Pro or Business Premium plan at no extra cost." For more on dots, see our Oct 12 explainer.

What to watch#

Every prompt we wrote keeps a person approving anything that goes out, changes the books or moves money. Of the builders, only Smith says outright that messages stay as drafts. The longest-running build is Simone's, about nine months by his account, and his retired-model failure is the best case for an error alert on any scheduled job. Gatte's dot is the newest here, about a week old when he posted.