https://www.youtube.com/watch?v=UXbAuWvUwUE Mike Russell (Creator Magic) went through the nearly 9-hour Grok Bot Galaxy Day One stream, pulled out five tips from the SpaceX AI team, and tested each one live in his own Grok Bot. Two of them cut token spend directly, one helps your bots write like you, and one keeps a growing bot team in sync without copy and paste. Tip 1: Turn Browsing Into an API Skill (Saves Tokens and Time) Source: Shubh, who works with founders at SpaceX AI, answering a too-expensive question - Idea: have the bot watch its own network requests while it does a browser task once, then call those requests directly next time How he tested it 1. Asked his chief of staff to spin up a new bot that uses its browser to find top things to do in Belfast on TripAdvisor 2. In the same prompt: capture the network requests, learn how to get this for any city, and record it as a skill 3. The new bot (TripAdvisor Scout) worked through the browser and its developer tools, then saved a reusable skill with hard rules, the API calls, and browser as a fallback 4. Test: use the skill, do not use the web browser, top 10 for Stockholm - Result: about 1 minute versus about 5 for the first browser run, roughly five times faster - Bonus: the bot updated its own skill with what unlocked the data (self-improving skill) Tip 2: Use Webhooks Instead of Every-15-Minute Routines Source: Shubh again: a 15-minute routine runs about 100 times a day, and most runs find nothing How he rebuilt a YouTube watcher 1. His chief made a YouTube watcher bot that checked his channel every 15 minutes 2. He told it: delete the 15-minute routine and set up a routine triggered by webhook instead 3. Open the routine, scroll to the bottom, copy the Post to URL and the key 4. In Zapier: trigger YouTube, New video in channel; action Webhooks POST to that URL 5. Header: Authorization with value Bearer plus your key; payload type JSON with title, description, and URL

https://www.youtube.com/watch?v=ff7om2bBLKM Nate Herk builds two email setups in Grok Bot. First, a bot that keeps his real inbox clean twice a day. Second, a bot with its own email address that wakes up the moment an email lands, so other agents and automations can hand it work by email. The first takes 5 to 10 minutes. The second takes a free AgentMail account and a webhook. What He Runs Today - Inbox bot (Nat): runs at 6:45 AM and 6:45 PM, labels every email, marks the noise as read, and sends him a short digest - Trader bot: has its own inbox; another AI agent does the research and emails it instructions like place this trade or sell this, and the email itself wakes the bot up - Key idea: a routine can fire on a schedule or when a webhook fires, and an incoming email can be that webhook Build 1: The Inbox Manager 1. Create a new bot and tell it its job: you are my email management agent, help me keep my inbox clean 2. Spell out your routing rules first: the labels you want and what belongs in each - Primary: real conversations with people - Urgent: bills, invoices, anything time-sensitive - Notifications: banks and services you signed up for - Outreach: sponsors, pitches, cold asks - The bot suggested adding Newsletters because he gets a lot; he said yes 3. Tell it what to do with each label: he has it mark everything except Primary and Urgent as read right away 4. Connect your email from the marketplace (Gmail, Outlook, etc.); you can add several accounts, just give each one a clear label 5. Ask for a routine twice a day with a digest sent to your phone (he uses a ClickUp DM): what came in, what is urgent, what to reply to 6. Optional: have it sign each digest (his ends with sent by Nat) so you know which bot sent it - His bot only labels and marks as read. It does not draft or send anything. Build 2: Give a Bot Its Own Email 1. Create a free AgentMail account and make a new inbox (default agentmail.to domain, or connect a custom one); the free plan includes a few inboxes

8. Weekly Time Tracking Analysis (chief of staff, weekly) - Reads the week’s time tracker files - Reports time spent per area (YouTube, shorts, community, and so on) and what to optimize next week Club Steal Sheet 1. Pick one routine from each group to start: competitor inspiration, prospect research, nightly improvement 2. Write each routine as an outcome plus rules (what to include, time window, what to skip) 3. Send results to a Notion page or Google Sheet instead of only chat so they pile up into a record 4. Run it once, check the output, and add a rule for anything it missed 5. Add a simple time tracker skill if you want tweet ideas and weekly reviews based on what you actually did 6. Keep the nightly improvement read-only first, then allow small fixes once you trust it Credit: Moritz | AI Systems — https://www.youtube.com/watch?v=0fyVaDintWQ

https://www.youtube.com/watch?v=zA1vAZr9u20 VelvetShark moved a daily OpenClaw workflow to Grok Bot and logged 67 real jobs over five days. This is a field report, not a hype reel: what finished on its own, what broke on day one, how each break got fixed, and the starter rules he would give anyone setting up today. The Scoreboard (His 5 Days) - 67 logged jobs - 40 finished with no help, 14 finished after he stepped in - 5 half-done, 8 failed or canceled - 50 of 67 needed no intervention at all - Weekly usage bar after all of it: about 29 percent, including setup What Actually Sold Him: The Brief 1. He typed a lazy prompt: Find me a place to fix jewelry 2. His manager bot turned it into a full brief for the research bot: his neighborhood, how long the shop has existed, reputation, what to avoid, and a ranking method 3. Those preferences (near me, established, owner-run, no chains, no mall kiosks) came from context the system already had about him, not from the prompt 4. Result: four ranked shops with hours, a three-line answer, and a recommended pick, for zero minutes of his time - Takeaway: the search is easy for any agent; the translation from lazy prompt to full brief is the real value The Setup That Worked 1. Day one plan failed: porting the old OpenClaw agent wholesale did not translate one-to-one 2. Rebuild: role-specific bots instead of one giant agent 3. Sharky is his COO / chief of staff bot; he talks mainly to Sharky 4. Seven specialists sit under Sharky and receive Sharky’s briefs 5. His line: agents know what agents need, so the manager writes better specialist briefs than he does 6. All bots share one cloud computer, one browser, and one system on a data center IP Failure 1: Data Center IP Blocks Logins - Symptom: the cloud computer could not sign in to X (Twitter); it only said the login was temporarily limited - Cause: the machine sits in a data center, so many sites treat its IP as a likely bot

https://www.youtube.com/watch?v=d7I82Ji45II Beginners in AI (Practical AI Agent series, part 4) builds SecondMed inside Grok Bot — a medical second-opinion helper that explains results, checks bills with Coupon, polls nutrition agent Lettuce, and preps doctor-visit questions from Google Calendar. Hard rule up front: this is not a doctor and not primary care. Create SecondMed 1. In Grok Bot, hit + for a new bot 2. Name it SecondMed 3. Upload a custom image if you want 4. Leave the description blank so the bot can fill it after instructions 5. Paste a clear role prompt: you are SecondMed on the health and wellness team with Wart and Lettuce; interpret charts/exams/images; act as a second opinion — not primary care What SecondMed Is For (And Not For) - Research, plain-English interpretation, appointment prep, second-opinion flags - Not your primary physician and not a replacement for human medical care - Useful for the online work you would already do yourself between visits Core Instruction Blocks To Paste 1. Role + team mates (Wart, Lettuce, and later agents) 2. Responsibilities: interpret charts, exams, images; invent extra assists that stay inside the medical-helper role 3. Flag when a human second opinion is warranted if a diagnosis or recommendation looks incorrect or incomplete 4. Coordinate with Coupon on medical bills: SecondMed sees the test; Coupon sees the billing code; ping you on over-bills or mismatches 5. Translate after-visit summaries, exams, and results into layman’s terms — clear, objective, never soften bad news 6. Flag values that are clinically in range but drifting toward an edge, and say what that might mean 7. Periodically poll Lettuce for diet/nutrition status so medical questions have context 8. Prep upcoming appointments: questions to ask, tests to inquire about, using history + teammate consults 9. Keep a running record; ping on red-flag trends without waiting for a new lab drop 10. Have Wart forward recall alerts so SecondMed can judge personal risk