Use Cases

21 real examples of businesses running work on AI, each with the source, how it went and a prompt to copy. Filter by your business and your tools.

For
Tool

Showing all 21

Evidence grade B

Operations

Give an always-on agent a revenue target and let it run the business

Nat Eliason's OpenClaw agent, Felix, builds and sells digital products around the clock and reports $177,417 in lifetime revenue.

How it went

In an interview, Felix reported $177,417 in lifetime revenue across products and services.

  • Revenue is self-reported from Nat's own dashboard.
  • Nat still sets strategy, makes judgment calls and corrects tone; Felix can't take calls and runs into context limits.
  • Distribution did much of the work: one analysis notes Nat's 100,000+ newsletter subscribers and 300,000+ X followers.
  • Nat put in long days early on, building the system around the agent.
View prompt
PromptWritten by our editors from the source
You are [name], operator of [company]. Target: $[X] in revenue by [date].

Every morning, write a five-line plan tied to the target.
Every night, post a revenue review to [channel]: sales, refunds, what worked, what to stop.

You may build and publish digital products priced under $[N] on [Stripe or Gumroad], post from [X account] up to [3] times a day, and answer customers at [agent email].

Ask me in [Discord] before refunds over $[amount], new spending over $[amount], legal terms or partnerships.
thenextnewthing.beehiiv.com
Evidence grade B

Content

Make launch and product animations with Claude in code, and budget for follow-ups

A creator's lab notebook shows Claude Opus 5.5 producing three studio-style animations in code, and how many follow-ups and hours each one really took.

How it went

Each piece took real iteration. The chrome loop took about 2 hours, 10 picture follow-ups and 12 transition checks, plus about 45 minutes and 5 follow-ups for sound.

  • The vague earbuds prompt came back with most of the clichés the directed prompt banned; the directed version took 15 follow-ups and about 90 minutes, plus 90 minutes and 12 follow-ups for sound.
  • The mascot took about 2 hours plus 70 minutes of sound; all 71 frame checks passed after 6 failures were fixed.
  • Claude can't hear: early mixes landed hits late and put energy where phone speakers play nothing.
  • A creator's own experiments, not client work.
View prompt
PromptWritten by our editors from the source
Make a 6-second seamless loop, 1080x1080 at 60 fps, as a single HTML file using three.js [pin the version].

Goal: [one sentence on what the viewer should feel or understand].
Timeline: 0–1.0s [action]; 1.0–2.5s [action]; 2.5–6.0s [action].
Use easing [cubic-bezier(0.22, 1, 0.36, 1)], a 60 ms stagger, and anticipation before every big move.
Ban: dark space backgrounds, glowing rings, bloom, floating particles and made-up taglines.

After building, render frames 0, 30, 60 and onward, check every transition against these timings, and list what's off by frame number before fixing it.
x.com
Evidence grade B

Support

Run customer support with a bot that asks before refunding

ChatPRD's support Bot works the Intercom and email queues every hour. In the week before the write-up, several customers asked where they could leave a review.

Grok BotStartupsApprove-to-send
How it went

Claire reports that in the past week, several customers asked where they could leave a review for the support they got, which had never happened before.

  • No resolution-rate or ticket-volume numbers were published.
  • Refund authority stays with a human by design.
  • At the time, Grok Bots couldn't join group chats.
View prompt
PromptWritten by our editors from the source
You are Holly, the support teammate for [product].

Every hour from [6am] to [9pm], sweep [Intercom] and [support inbox].
Resolve anything the playbook at [link] covers, and reply in our voice: short, warm, specific.

You may request refunds but never issue them. For each one, send me an approval card with the customer, order, amount and reason.
Escalate billing disputes, security, legal questions and angry customers to me with a two-line summary.

Every Monday, review last week's tickets and propose three additions to the playbook or help docs.
chatprd.ai
Evidence grade C

Support

Automate order-status and return tickets on a Shopify store

Footwear brand Orthofeet automated 56% of support tickets in under two months with Gorgias's AI Agent, beating its goal of 30% in 30 days.

How it went

Gorgias reports 56% automation in under two months, first responses down to 35 seconds for email and 13 seconds for chat, and double-digit yearly growth without adding headcount.

  • Vendor case study.
  • A third-party review of Gorgias's case studies puts typical automation between 26% and 56%; Orthofeet sits at the top.
  • Model the AI Agent's cost against your ticket volume before switching it on.
  • Text channels only; phone volume is not automated.
View prompt
PromptWritten by our editors from the source
Audit our last 500 support tickets. Group them by intent: where's my order, returns, exchanges, sizing, product questions, other.

Show volume and average handle time for each, and mark which ones can be answered from order data and our policy pages alone.

For the top three, draft guidance in our brand voice with clear hand-off rules: health questions, angry customers and orders over $[amount] go to a person.
gorgias.com
Evidence grade C

Content

Run a weekly local newsletter with a six-Bot team

The Arlington Bagel, reported at about 6,000 weekly readers, runs research, compilation, editing and email on six Grok Bots that replaced an intern.

How it went

Per the write-up, six Bots now cover research, content compilation, editing and email correspondence, replacing a human intern.

  • No revenue, open-rate or time-saved figures were given for the newsletter.
  • This is a secondhand write-up by Stork.ai, not Billy's own account.
  • Greg Isenberg, quoted in the write-up, advises keeping Grok Bot on one project; running several businesses at once lets context bleed and burns through usage.
View prompt
PromptWritten by our editors from the source
You are Chief of Staff for [newsletter], a weekly newsletter about [topic or town] for [audience].

Read our Notion, Slack and Gmail. Write a one-page audit: what we publish, each step and how long it takes, and where revenue comes from.

Recommend the two or three Bots that would save the most time or earn the most first, with a one-paragraph job description for each.

Don't create Bots, send email or change anything until I approve.
stork.ai
Evidence grade C

Marketing

Run SEO with Claude doing the work and Search Console data doing the thinking

ChatSEO's co-founder says pruning and retargeting pages took a 50-page blog from 1,500 to 15,000 monthly clicks in three months, and documents the times Claude broke live pages.

How it went

He reports going from 1,500 to 15,000 monthly clicks in three months on a 50-page blog without publishing a new article, by cutting dead pages and pointing pages at queries Google already showed them for.

  • Self-reported, and he doesn't say how much of the gain came from AI.
  • Failures: Claude rewrote city pages and dropped their ranking tables, found live during a podcast recording; another edit deleted the key section of a comparison page.
  • On a fresh domain chasing competitive terms, he says the same process produces pages that rank around 40th.
  • Promotional: the course routes data through his product. A free route uses Search Console exports and Google Suggest.
View prompt
PromptWritten by our editors from the source
Using my Search Console data, list every page with zero clicks and fewer than 50 impressions in the last 90 days, excluding pages published in the last 3 months.

For each page, recommend delete, merge (and into which page), or rewrite, with one line of reasoning from the data.

Safeguard before any live edit: list every table, form, call to action and embed on the current page, confirm each one survives in your edited version, and stop to show me the difference if anything is missing.
x.com
Evidence grade B

Finance

Collect tax documents from Gmail and build deduction spreadsheets

A consultant built two Claude Code skills: one finds tax documents in Gmail and tracks what's missing, the other compiles deductions into spreadsheets ready for tax software.

How it went

He found Claude Code surprisingly good at the grunt work: sorting transactions, filtering by category, parsing CSV exports and assembling spreadsheets.

  • His guide says many people shouldn't do their own taxes this way, and explains when to get help.
  • Never paste Social Security or bank account numbers into a session.
  • No time-saved figure is given.
View prompt
PromptWritten by our editors from the source
Search my Gmail for [tax year] tax documents: 1099s, 1098s, W-2s, K-1s and receipts from [vendors].

Save attachments to ./tax/[year]/inbox and list each one in ./tax/[year]/documents.csv with payer, form type and amounts.

Compare against last year's list in ./tax/[last year]/documents.csv and tell me which forms are still missing.

Redact Social Security and full account numbers, and never store them.
claudeblattman.com
Evidence grade B

Content

Curate newsletter sources and first drafts with a writer Bot

After two days of testing, Ryan Carr of a done-for-you newsletter agency found Grok Bot's drafts matched his cadence and its curation found better sources than vanilla Claude.

How it went

After two days he reported drafts close to his cadence that still needed a human pass, research that returned usable patterns instead of a dump of posts, and stronger large-scale curation than vanilla Claude.

  • Early impressions after 48 hours, not a client result.
  • He isn't switching yet: his projects and prompts live in Claude Cowork and Claude Code, with no clean migration path.
  • Beta gaps he hit: no one-click 'open in Google Drive' and no per-client project knowledge sidebar.
View prompt
PromptWritten by our editors from the source
You're my newsletter researcher for [publication] (audience: [who]).

Every [Tuesday], collect 25 candidate stories from [sources] published in the last 7 days. Skip anything the big outlets already covered unless there's a new angle.
For each, give a one-line summary, why our readers care, and the link.

Then hand the top 8 to the Newsletter Writer Bot to draft the issue in our voice, using [last 3 issues] as examples. Park the draft for my edit.
moodboard.beehiiv.com
Evidence grade B

Marketing

Draft LinkedIn and X posts in your own voice, ready to approve

Debbie O'Brien set up one Bot per channel on launch day. Her LinkedIn Bot pulled her recent posts to learn how she writes.

How it went

Her takeaways: name the Bot after the job, give it a real job, and don't create a Bot for every subtask.

  • A setup walkthrough, not a results report: no engagement numbers.
  • The walkthrough doesn't cover approvals; keep posting behind yours.
View prompt
PromptWritten by our editors from the source
You're my LinkedIn Bot.

First, read my last 30 posts and write a six-line style guide: length, hooks, tone, formatting, emoji and calls to action. Wait for my OK.

Then, twice a week, draft two posts from [my blog, changelog or Notion notes] and park them for my approval.

Never publish, comment or message anyone without my approval.
debbie.codes
Evidence grade B

Sales

Get a research brief on every prospect 10 minutes before the call

Annika Helendi copied Lindy's sales-call prep template: it researches each discovery-call contact and emails her a brief 10 minutes before the meeting. It worked within minutes.

How it went

The call-prep template and the notetaker worked out of the box. The content system did not: constant errors burned through her monthly credits, so it stayed on Make.

  • Credit costs made her avoid experimenting, including with the YouTube agent.
  • She reported 400 free credits without premium actions and $50 a month for 5,000 credits; the review is undated, so check current pricing.
  • Her rule of thumb: fixed, step-by-step workflows belong in Make; simple agent tasks in Lindy.
  • The review includes her Lindy referral link.
View prompt
PromptWritten by our editors from the source
When a calendar event with "discovery" in the title is 10 minutes away, research each external attendee: LinkedIn, company site, Crunchbase and recent news.

Email me a brief under 150 words: role, company stage, likely priorities, three sharp questions, and one thing we have in common.

If you can't verify something, leave it out.
annikahelendi.substack.com
Evidence grade B

Product

Ship a product in an afternoon with an AI code editor, then sell sponsor slots

Pieter Levels prompted Cursor into a browser flight simulator in about three hours and reported $1M in annualized revenue from in-game ads 17 days later.

How it went

Levels reported $87,000 in monthly revenue, a $1M annual run rate, 17 days after launch.

  • A run rate, not $1M collected.
  • Coverage pointed out that his roughly 623,000 X followers drove much of the demand.
  • A 2026 recap reports the game's revenue later fell to zero.
View prompt
PromptWritten by our editors from the source
Build [product] as a single-page web app in plain HTML and JavaScript, with no framework.

Start with the smallest version someone can use in 30 seconds. Then list the next five improvements and wait for me to pick.

Add a "sponsor slot" component that reads advertisers from a JSON file, so I can sell placements without touching the code.
404media.co
Evidence grade B

Operations

Triage six inboxes every hour and only get pinged for what matters

A chief-of-staff Bot sweeps six email accounts, Slack and calendars every hour, archives the noise and sends only high-signal alerts plus a morning brief.

How it went

It flags things like a spike in data transfer on her website, noting it can wait until after her meeting. Routines needed more explicit setup than in OpenClaw, but she calls them rock-solid once set.

  • She still writes high-stakes emails herself.
  • Grok's default writing read as clipped until it learned from her sent mail.
View prompt
PromptWritten by our editors from the source
You're my chief of staff.

Every hour from 6am to 9pm, sweep [inboxes], [Slack workspaces] and [calendars].
Archive newsletters, receipts and notifications. Leave anything I can deal with later unread.
Message me only about: money, unhappy customers, deadlines in the next 48 hours, and anyone on [VIP list].

At 7am, send a five-line brief: what's new, what's due, what needs me.

Before you draft any reply, read my last 200 sent emails and show me a short style guide of how I write.
chatprd.ai
Evidence grade C

Marketing

Find email flows that are double-messaging your customers

April Polk of JaeLuxe Shoetique, an early Muse tester, had it find and shut down duplicate Klaviyo flows that were emailing customers twice.

How it went

Per Meta's launch post, the duplicate flows were found and turned off. No figures on unsubscribes or revenue were shared.

  • A Meta-selected tester quote for a launch: the best case, not an average.
  • Muse is available in the US and Canada only.
  • Changes require approval by default; keep it that way for anything customer-facing.
View prompt
PromptWritten by our editors from the source
Review every active flow and campaign in Klaviyo.

List any pair of flows that can email the same person within 48 hours, with each one's trigger, audience and sends in the last 30 days.

Recommend which to pause or merge, and why. Don't change anything until I approve.
about.fb.com
Evidence grade C

Marketing

Generate upload-ready Google Ads variations with your brand rules built in

A growth marketer who had never coded built a Claude Code workflow and a Figma plugin that cut ad-variation production from about 30 minutes to about 30 seconds.

How it went

Anthropic reports ad-variation production dropped from about 30 minutes to about 30 seconds per variant.

  • Vendor case study: Anthropic writing about its own product.
  • A third-party write-up puts a full campaign refresh at about two hours down to 15 minutes.
  • The workflow enforces Google's character limits instead of leaving them to chance.
View prompt
PromptWritten by our editors from the source
Build me a /rsa command for Google Ads responsive search ads.

Ask me for: campaign name, landing page, top 10 keywords, and the current headlines and descriptions with their performance.

Write 15 headlines (30 characters max) and 4 descriptions (90 characters max) that follow ./skills/brand-voice.md. Flag any claim not backed by ./product-facts.md.

Show character counts, refuse to output any line over the limit, and export a CSV I can upload to Google Ads Editor.
claude.com
Evidence grade C

Content

Let Claude Code hand X research to Grok Bot mid-task

Riley Brown demoed Claude Code calling Grok Bot through a webhook skill whenever it needs current X posts, then finishing the original task with the answer.

How it went

A demo video plus a setup guide, with no business metrics. One reply says they set it up in about 30 minutes.

  • Promotional: the author's company, Agent Native, publishes the guide, which requires a sign-in.
  • Several replies argue the reverse works better: Grok Bot orchestrating, Claude Code doing the work.
  • Open questions went unanswered: how accurate the research is, and whether the Bot can act on the X account or only read it.
View prompt
PromptAdapted from the source
Skill: grokbot-research

When you need current X posts or discussion about a topic, POST the question to [webhook URL], wait for the reply, cite the post links it returns, then continue the original task.

Test task: Find the 10 most-discussed X posts this week about [your niche], summarize the angles, then draft a landing-page section that answers the top objections.
x.com
Evidence grade C

Finance

Total photographed time cards for payroll

An Iowa grocer who works about 65 hours a week had Muse add up photos of employee time cards, cutting at least 15 minutes from payroll.

How it went

Per Meta's launch post, it cut at least 15 minutes from each payroll run.

  • A Meta-selected tester quote, with no detail on accuracy checks.
  • Check totals before paying anyone; handwriting and missed punches trip up any reader.
  • US and Canada only.
View prompt
PromptWritten by our editors from the source
Here are photos of this week's time cards.

Extract each employee's in and out times into a table, and total regular and overtime hours (overtime after [40] hours a week).

Flag any card that's hard to read, missing a punch, or over [10] hours in a day.

Don't send anything to payroll. Give me the table to check.
about.fb.com
Evidence grade D

Marketing

Audit your Google Business Profile against the businesses outranking you

A local SEO marketer published 20 Grok Bot prompts that browse Google Maps, pull competitors' categories and attributes into spreadsheets, and run a 12-week audit plan.

How it went

No measured results. His claims are anecdotal: clients who added one secondary category started showing up for new searches the following week.

  • Promotional: the article ends with a pitch for his agency.
  • Everything depends on the Bot reading Maps and profile pages correctly; check the data before you change your profile.
  • Published in August and re-shared on October 14, 2026.
View prompt
PromptWritten by our editors from the source
Use the business brief I gave you.

In your browser, search Google Maps for "[service] in [city]" for each of these keywords: [keyword 1], [keyword 2], [keyword 3].
For each search, record the businesses in the top three map results and their positions. Open each competitor's Business Profile and note its primary and secondary categories, rating and review count.

Build a spreadsheet with one tab per keyword. Then list the categories competitors use that my profile is missing, ranked by how many top competitors use them.
x.com
Evidence grade D

Support

Draft support replies inside your helpdesk with a self-hosted agent

A small-business owner wired Hermes Agent into Front with a write-only token so it drafts replies for his team. The goal is 80% less email writing; it is still a work in progress.

How it went

It reads the thread and writes a draft, but he said looking up answers in their software was still unreliable and the project was way harder than expected. The 80% figure is his goal, not a result.

  • A short, self-reported update, quoted in a third-party review; we couldn't load the original post ourselves.
  • Hermes is not a helpdesk product: by default its messaging gateway rejects unknown users, and its memory is shared across chats on a profile.
  • Keeping it behind your team, drafting only, is the safe shape.
View prompt
PromptWritten by our editors from the source
For each new thread in [helpdesk], read the whole conversation, then search [FAQ built from our last 500 sent replies] and [help docs].

Write a draft reply in our voice. If you can't find a supported answer, leave an internal note saying what's missing instead of guessing.

Never send. Drafts only.
x.com
Evidence grade D

Operations

Give your Grok Bot team Claude and ChatGPT models by installing their CLIs

A coach had his Chief of Staff Bot install the Claude Code and Codex command-line tools on its own computer, then routed writing to Claude Opus 5.5 and coding to another model.

How it went

He says it works extremely well, but gives no metrics or before-and-after comparison.

  • Replies raise unverified concerns: whether this fits Anthropic's terms, and whether signing in from a data-center IP risks a ban.
  • One user found their CLAUDE.md settings and memories didn't carry over to the Bot's computer.
  • Others say Grok Bot already routes some tasks to models like Opus 5.5 natively, so this may matter less over time.
View prompt
PromptWritten by our editors from the source
Open a terminal on your computer and install the Claude Code and Codex command-line tools. Tell me when each one needs me to sign in, and wait.

After I'm signed in, update every Bot's instructions and our shared memory with this rule: use Claude Opus for writing tasks and [your chosen coding model] for coding and computer-use tasks.

Show me the exact text you changed.
x.com
Evidence grade D

Operations

Onboard your first Grok Bot like a new hire, with a probation period

A widely shared guide recommends a five-field job description, a boring first task, a dry run, and promotion up an autonomy ladder only on evidence.

How it went

It is a framework, not a reported result. Its strongest part is the failure list: archiving your accountant's email thread, trusting a source it should have flagged, or repeating an allowed action four hundred times after something upstream changed.

  • Promotion gate: five clean runs in a row, a tested rollback, and no unresolved side effects.
  • Bots on one account share files, browser sessions and logins; as the guide puts it, names are "a visual boundary, not a security boundary."
  • Anonymous author with very high reach; no case study.
View prompt
PromptWritten by our editors from the source
Job: [one sentence].
Owns: [the outcome you're responsible for].
Inputs: [the brief, approved sources, past work].
May: [search, read, compare, organize, draft].
Must ask before: [contacting anyone, buying, publishing, deleting, moving money].
Done when: [every claim has a source, conflicts are listed, gaps are written down].

Before doing anything, don't execute. Walk me through each step you'd take, every tool you'd open, and every judgment call you're unsure about.
If sources conflict or anything is ambiguous, stop and ask me.
x.com
Evidence grade D

Operations

Send your whole Bot team to find one improvement each

A one-line Chief of Staff prompt sends every Bot off to improve something on its own. Users say it finds good gaps, and can eat a big share of a week's usage.

How it went

The post reports no results, and replies are mixed. One user said it worked well but left hours of work to review.

  • Usage: one reply said it used a lot of their limit; another joked it cost 40% of their usage.
  • Collisions: one Chief of Staff handed out 12 jobs at once, every Bot tried to use the same computer, and it got stuck.
  • One user runs it every three weeks.
  • The 'nothing that needs my permission' clause carries a lot of weight.
View prompt
PromptWritten by our editors from the source
Get all the Bots together. Give each one a single improvement to investigate in its own area, no more than [3] at a time so they don't fight over the shared computer.

Don't send, post, buy, delete or change permissions.

For each improvement, send me a one-paragraph proposal with the expected benefit and how much usage it will take, and wait for my go-ahead before doing it.
x.com

Grades: Evidence grade A Verified Evidence grade B First-hand Evidence grade C Vendor/secondhand Evidence grade D Anecdotal How we grade