Real things people did with AI agents, each with a source and a prompt you can copy. Summaries and prompts are written by Milo; results are as reported by the source.
A builder split finding, planning and building across two scheduled Grok bots and a Claude Code factory in GitHub Actions. The Scout files ideas as Goal issues; the Builder turns an approved idea into a spec and tickets; the factory opens draft PRs and may rework a ticket at most twice. He stays in the loop at two points: approving the spec and merging the PR. One public run went from Goal #108 to a merged pull request (#118).
Set up a small software factory on repo [REPO] that uses labels as the only hand-off. One scheduled agent (the Scout) may only find ideas and open Goal issues; it must not plan or write code. A second scheduled agent (the Builder) may only turn a Goal I have approved into a spec and tickets. A factory may open draft pull requests and rework a ticket at most twice, then stop. Never merge, deploy, force-push or mark a PR ready. I am the only one who approves a spec and who merges. Write the label state machine and the 'never' list first, then the prompts.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A Meta Muse user dictated date, time and party size; Muse called the restaurant, booked the table and shared a transcript. When the booking didn't show online, Muse called back and the host confirmed it.
Call [RESTAURANT] and book a table for [N] people on [DATE] at [TIME] under the name [NAME]. Mention it's a [OCCASION]. Ask for a confirmation name or number, and send me the call transcript. Two days before, call again to confirm the booking is still on.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A personal AI assistant went through a year of card statements, found a duplicate DashPass subscription, and negotiated an internet bill down in the provider's support chat while the owner approved each step. It also miscounted declined charges, which the author corrected.
Read my card transactions from the last 12 months (posted charges only, ignore declined ones). List every recurring charge with amount, start date and whether it changed price. Flag duplicates and price increases, then suggest which to cancel or negotiate. Don't cancel or change anything without asking me.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
Devin ordered two iPhones from Apple with Affirm financing and trade-ins, then refused delivery of one and lost track of the refunds. He says Grok Bot, watching his email and accounts, spotted that Affirm had collected $980 for the phone he sent back while Apple never refunded it, and pulled together the order numbers, tracking and what to say. He called Apple himself, and Apple confirmed the refund was issued.
I returned or refused [ITEM] from [STORE] on [DATE]. It was paid with [CARD OR FINANCING, e.g. Affirm] and may include a trade-in. Search my email and the [STORE] and [LENDER] accounts and build a timeline: what I was charged, what was refunded, what the trade-in credited, and what is still being collected. Flag any money going out for an item I no longer have. For each problem, list the order number, return tracking, dates and amount, plus a short script I can use when I call. Do not contact anyone or change anything without asking me.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
Aaron says OpenClaw found his county's appeal form, researched comparable home sales, filled the form in and submitted it. A week later the county agreed and lowered his property taxes.
I need to appeal my property tax assessment for [ADDRESS] in [COUNTY]. Find the official appeal form and the deadline. Research recent comparable sales nearby and summarize them in a table (address, sale date, price, size, how it compares). Draft the appeal with those comparables, then show me the completed form to review. Do not submit anything until I say so, and do not invent sale prices or sources.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A founder has Claude posting faceless motion-graphics explainer videos for his app across 9 Instagram, YouTube and Pinterest accounts. After a week, views went from about 100 to 300-600 per video, he reports.
Here is my app: [LINK]. Write 7 short explainer video scripts (under 40 seconds each), each showing one problem the app solves. For each, give a hook for the first 2 seconds, on-screen text, a caption, and hashtags for Instagram Reels, YouTube Shorts and Pinterest.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A user had their AI agent handle three separate support chats with AT&T about three $35 charges from one store visit; all $105 was credited without them making a call.
I was charged [AMOUNT] by [PROVIDER] for [CHARGES, DATES]. Open their support chat, explain politely why these charges are wrong, and ask for a full credit. Keep a log of what each agent promises, and check the bill afterwards to confirm the credit actually posted.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A gardener instrumented a 2 m² vegetable bed with soil probes, a weather station and a camera, then scheduled a morning agent to read the history and the forecast and adjust watering. A fixed rules engine kept being wrong when context mattered. The agent may make at most one change per zone per run, keep minutes between 1 and 15, and must log what it expects so it can check the next day. An old Android phone taps the official Rain Bird app to apply the change. As of 5 November 2026 the garden had a little over four million readings.
Every morning, review yesterday's [SENSOR READINGS], the photos, and the [WEATHER FORECAST]. Decide whether to change watering for each zone. Hard limits: at most one change per zone per run; change duration by at most 20%; keep minutes between 1 and 15; never turn watering off unless at least 5 mm of rain is forecast in the next 24 hours. Write down what you changed and what you expect to see in the soil tomorrow, then check that expectation on the next run. If you are unsure, do nothing and tell me why.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A builder set up scheduled Grok Bot routines with Gmail and Calendar connectors: action threads stay, everything else is labelled and archived, and replies are saved as drafts only. He estimates 1.3 to 1.9 hours saved a week.
Twice a day, go through my new Gmail. Keep only threads that need a reply or a payment in the inbox; label and archive newsletters, receipts and updates. Never send, delete or trash anything. Then send me a short digest: what needs me, what you filed, and my calendar for the next two weeks.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A small-business owner had an AI assistant triage email through six ordered questions, propose meeting times and draft routine replies, with hard rules: never send, never delete, never commit to prices or dates.
Triage my inbox. For each email, stop at the first yes: is it noise (archive), already handled (archive), urgent today (label @Urgent), a meeting request (suggest 3 times from my calendar), or a routine reply with known facts (draft it)? Everything else: label @Action Required with a one-line summary. Never send, delete, or agree to any price or date.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A Meta Muse user has the agent crawl a slow meal-delivery web app once a week and surface the options that match their dietary preferences, so they don't get stuck with unappetizing defaults. The same person also has Muse search Craigslist and Facebook Marketplace daily for apartments that fit a long list of requirements, and only pings them when something worthwhile appears. They have not asked it to email landlords.
Once a week, open my meal-delivery account at [URL], go through next week's menu, and list the dishes that fit these preferences: [DIETARY RULES]. Skip anything I marked as a skip last time: [DISLIKES]. Give me a short shortlist with why each one fits, and wait for me to confirm before you place or save an order. Do not pay for anything.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A consultant had Claude Code turn a plain-language target profile into Clay searches, research each contact individually, and hand back a finished draft in his own voice for every lead.
Target: [INDUSTRY, SIZE, LOCATION, ROLE]. For each matching contact, research their company and recent public activity, then draft a 4-sentence email in my voice (examples: [PASTE 2 EMAILS]) that references one specific, true detail. Put everything in a table for me to review.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A GTM founder scraped, enriched, personalized and launched an outbound campaign by prompting Claude Code with a GTM CLI in a few exchanges, cutting setup from hours to about forty minutes. The first 180 emails produced eight positive opportunities, he reports.
Build an outbound list for this ideal customer: [ICP]. Find 200 matching companies and the right decision maker at each, verify emails, and write a short personalized first line for each based on something true and recent about the company. Output a CSV. Don't send anything.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A founder uses Claude Code beyond coding: SEO audits of structured data across 25 language versions, batch blog translations, and performance debugging that found an empty server-rendered page.
Audit my site [URL] for SEO: check titles, meta descriptions, canonical tags, structured data validity and hreflang. Also check that the server-rendered HTML actually contains the page content. Give me a prioritized list of fixes with the exact file to change.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A developer asked Claude Code to audit a repo for secrets before going public. It missed a Contentful management token in .claude/mcp.json, a file Claude Code had created, which had been committed into git history.
Audit this repo for secrets before I make it public. Scan the full git history (not just current files), including hidden folders like .claude/, .cursor/ and .vscode/, plus .env files and CI configs. List every credential with file, commit and type, then tell me which to rotate and what to add to .gitignore.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A developer built a Telegram bot that transcribes voice memos and uses the Claude Agent SDK so you can later ask things like 'what did my manager say about the deadline?' and get answers from your transcripts.
Here is a folder of my meeting and voice-memo transcripts: [FILES]. Answer: [QUESTION]. Quote the exact lines you used, with the file name and date, and say clearly if the answer isn't in the transcripts.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
The host of a tech show built three OpenClaw skills that find trending moments on X, scan long YouTube videos, and dig up 'on this day' moments from 15 years of episodes to make social clips.
Watch this long video: [URL]. Find the 3 moments most likely to work as standalone 30-60 second clips (strong opinion, surprise, or a clear takeaway). For each, give start/end timestamps, a hook headline, and caption text. Explain briefly why each one should work.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A new dad built Babylog, a feeding, diaper and nap tracker, entirely by chatting with an OpenClaw agent over Telegram. The agent handled architecture, code, Cloudflare Workers deployment and CI/CD.
Build me a simple mobile-friendly web app to log [THINGS TO TRACK] with timestamps, plus a daily summary view. Keep it to one repo, deploy it to [HOST], and set up automatic deploys on every push. Explain each step in plain English before you run it.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A developer gave Claude Code CLI tools over a library of 100 non-fiction books and let it explore. It produced more interesting thematic 'trails' between books than his hand-built LLM pipeline.
Here are my notes and highlights from these books: [FILES]. Find 5 surprising connections between ideas in different books. For each, quote the passages, explain the link in two sentences, and suggest one question it raises.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.
A developer had Claude Code build his wife a local dashboard with a pomodoro timer, task manager and notes, replacing three websites. Data stays on her machine, and it took a 5-minute spec.
Build a single local web page that combines [TOOL 1], [TOOL 2] and [TOOL 3]. Store all data locally in the browser, no accounts and no external services. Give me one command to start it.
A generic template written by Milo for this use case, not the original user's prompt. Swap the [placeholders] before you run it.