Reported, not yet confirmed.
Five weekly AI builds you can copy this week
Steal one recurring job from this Thursday showcase: client recaps, NDA filing, morning loose ends, Canvas grading, or a family meal plan.
Pick one recurring job this week and copy the shape below. Each build is someone else's real weekly (or daily) run, cited, with a starter prompt you can paste. Times and results are self-reported. We have not reproduced these runs on this box.
1. Weekly client recap (Grok Bot)#
Derek (@DerekatQuango), co-owner and account manager, had Grok Bot draft the weekly recap email to one client.
- Task: write that recap by hand each week.
- Tool: Grok Bot (already doing most tasks for that account).
- Result: "These emails used to take me 30-45 minutes each week." Now he cleans up the bot draft in "closer to 7-8 minutes a week," and says he "saved ~2 hours a month."
How to copy it
- Point a Bot at one client where it already sees tasks, mail, or notes for that account.
- Add a weekly routine that drafts the recap to you only. Do not send to the client without your review.
- Keep a short checklist of what must appear (wins, open items, next week). Edit the draft against that list.
- After two weeks, compare minutes to your old hand-written version.
Our starter prompt:
Every week, draft a client recap email for [CLIENT] using only what you already
know from this account's tasks, mail, and notes. Structure: (1) what shipped,
(2) open items and owners, (3) risks, (4) proposed next week. Mark anything you
are unsure about as [NEED CONFIRM]. Send the draft only to me. Do not email the
client.
Gotcha: this only works when the Bot already has the account's history. A cold Bot with no context will invent filler.
2. Weekly NDA filing (ChatGPT)#
Sunnyvale published a ChatGPT scheduled sweep that finds executed NDAs and MNDAs in Gmail and files missing copies on the HubSpot company record and the matching Google Drive customer folder. The article opener says the example "runs on my ChatGPT account weekly."
- Task: hunt the inbox for signed NDAs, then upload without duplicates.
- Tool: ChatGPT scheduled prompts with Gmail, HubSpot, and Google Drive.
- Result: an unattended weekly pass with a reconciliation report and an exceptions list. No minutes figure was published.
How to copy it
- Ask ChatGPT to draft the sweep in plain language (author starter below).
- Keep the safety rules: drafts and redlines are not executed agreements; never create a HubSpot company; never guess a Drive folder; never delete or overwrite.
- Test three cases: already filed (no write), missing from HubSpot only, ambiguous company (exception).
- Schedule weekly with a rolling 21-day Gmail lookback so prior finds can reappear and duplicate checks stay mandatory.
Our starter, adapted from the author's request (not their exact words):
Write me a prompt for a weekly scheduled sweep of my email box.
Look for NDA and MNDA agreements. Then look in HubSpot for the company
object and the customer folder, and find if the NDA is already there.
If it is not in either place, add it to both. Do not create duplicates.
Gotcha: this is filing, not legal review. Ambiguous companies and folders stay as exceptions.
3. Daily Loose Ends (Claude Code)#
Samuel Appentsen (@champagnekofi) runs a Claude Code morning routine that emails one "Daily Loose Ends" packet.
- Task: open email, Slack, calendar, and half-finished sessions to rebuild what still needs you.
- Tool: Claude Code on a morning schedule, with email and Slack.
- Result: "I skim it in five minutes, tick things off, snooze some, and draft replies straight from it." It opens with top three, today's calendar, then every reply owed with days waiting.
How to copy it
- Connect email (read and mail-to-self), Slack read, and the project folders or sessions that hold half-done work.
- Define a loose end: unreplied mail older than N days, Slack asks aimed at you, open decisions, unfinished local work.
- Shape the email: top three, calendar, reply queue with age. Disable outbound sends except the one email to yourself.
- Schedule before you open Slack. Trim false positives in week one.
Our starter prompt:
Every morning, build my Daily Loose Ends email and send it only to me.
Sources: my email, Slack, today's calendar, and my active project folders.
Include: (1) top three priorities for today, (2) today's calendar,
(3) every reply I owe in email or Slack with how many days it has waited,
(4) projects that look halfway done.
Do not send any reply to other people. Leave anything ambiguous in Exceptions.
Gotcha: keep each person's packet private. The author still ticks, snoozes, and drafts; the routine surfaces, it does not clear the queue.
4. Weekly Canvas grading assist (Claude)#
Daniel Lakens uses Claude to write R scripts against the Canvas API for weekly assignments across about 160 students.
- Task: grade weekly Canvas assignments with about 20 clicks per student in a slow UI, and hand out completion points without checking uploads.
- Tool: AI-written R scripts against the Canvas API (he uses Claude, including in VS Code, on a Claude 20x Max plan), plus Claude to check completion uploads.
- Result: "approximately 2 hours per week, for 8 weeks," with higher accuracy on low-stakes work he used to mark leniently.
How to copy it
- Have Claude write an R script that downloads assignments into a spreadsheet (questions and answers), not grade blindly in the LMS UI.
- Grade in the sheet against your rubric. Use Claude separately to flag completion-only uploads that look fake.
- Upload scores with a second R script. Keep final grades and high-stakes exams human.
- Document what share of the final grade these assignments are.
Our starter prompt:
Write an R script that uses the Canvas API to download this week's assignment
submissions for course [ID] into a spreadsheet with columns: student, question,
answer text, attachment names. Do not assign grades. Add a second script stub
that uploads scores from a graded sheet. Never send student data to a model
that is not covered by our institution's agreement. Flag, do not auto-fail,
ambiguous completion checks for my review.
Gotcha (student data): Lakens notes the AI-graded assignments are "a small percentage of the final grade, unlikely to determine who passes a course." Keep student work inside approved tools. Do not paste identifiable submissions into a consumer chat without your school's policy.
5. Weekly meal plan and groceries (Grok Bot + ChatGPT)#
Lyndall Schreiner runs a custom Grok Bot on a Markdown recipe library, then pastes the shopping list into ChatGPT to fill Instacart.
- Task: plan 15 family meals a week and shop after years of spreadsheets and a vibe-coded app.
- Tool: Grok Bot for plan and list; ChatGPT for the Instacart cart. Public template: https://x.ai/bot/dHdfbaNGz_LpSdKUhptdd
- Result: "Grok now plans 15 meals for my family each week and does my grocery shopping." No minutes figure.
How to copy it
- Fork or start from her public template, or give your Bot a folder of Markdown recipes.
- Weekly: ask for a printable plan plus a shopping list grouped by store section.
- Paste the list into ChatGPT (or your cart tool) only after you approve the plan.
- Check the checkout before anything is purchased.
Our starter prompt:
Using only recipes in my library, plan 15 meals for the next 7 days for a
family of [N]. Output: (1) a one-page meal plan printout, (2) a shopping list
grouped by produce, dairy, pantry, meat, other. Do not place any store order.
Do not subscribe me to anything. Wait for my OK before building a cart.
Gotcha: check the checkout before it buys. A Bot that can shop can also create a recurring order you did not mean to keep.
What to watch#
Watch for authors who publish the full skill file or routine JSON, not only the story. Prefer builds with before-and-after minutes (Derek, Lakens) when you pick what to try first. If you ship a Grok Bot version of Sunnyvale's NDA sweep or champagnekofi's Loose Ends, keep create-company, delete, and outbound reply actions behind approval.