Source: raw/Use_Grok_Bot_Better_Than_99%_of_People_Full_Course.md

Creator: Greg Isenberg (The Startup Ideas Podcast) with Ross Mike · URL: https://www.youtube.com/watch?v=1cWZ7Ekff6E · Duration: not stated in the raw file · Platform: YouTube

Ross Mike uses Grok Bot as the front end for nearly everything he does, and uses it mostly to orchestrate other tools rather than to do the work itself. Engineering goes to Claude Code sessions that Grok Bot starts; design goes through the Paper MCP; content ideas come from a twice-daily X scan; finance runs through Plaid-connected advisor bots. His case for Grok Bot over using Claude Code directly is the phone UX plus one agent that holds business, finance and engineering context together. Both hosts say they are not affiliated with Grok Bot or Paper.

Transcript normalizations: “Grockbot”, “Grogbot”, “Grockpot” → Grok Bot; “cloud code” → Claude Code; “Opus 55” → Opus 5.5; “codeex” → Codex; “Mobin” → Mobbin; “11 laps” → ElevenLabs. “Obliterated model” is kept as he says it (the usual term for a model with refusals removed is “abliterated”

Key Takeaways

  • Context decides when a project gets its own bot. He keeps one general engineering bot for small ideas (a screenshot extension took “two prompts”). A project gets its own bot when he takes it seriously or wants to ship it. When it stops mattering he deletes the bot (“right click and I delete”). Some bots are permanent: one manages his machine fleet, one handles local models and fine-tuning, plus his content bots.
  • Hand ideas to the engineering bot overnight. “Make this, give me something that works, push to my GitHub and I’ll check it later.” He woke up to an AI-governance app he has not yet run.
  • Ask for a review before a build. He gave a bot a plan PDF a friend made in Claude Code with only “Review this.” It replied “Good direction, but I wouldn’t approve the build plan unchanged,” listed risks and proposed a scope. He pasted his friend’s reply back in, approved the revised plan himself, then said “build it out.”
  • Grok Bot orchestrates Claude Code. He picks Claude Code over Codex because “Opus 5.5 is the better model” and because of the subscription subsidy. He says the 8,000 to $12,000 worth of API token credits”; that is his estimate, not an Anthropic figure. His prompt is short: “Build out everything. Use Claude Code with ultra code Opus 5.5. If you need multiple sessions go ahead /go.” He reports five out of five working PRs after it ran for hours. (“Ultra code” is as transcribed.)
  • Two ways to connect Grok Bot to Claude Code. The easy way: “every bot has a computer,” so tell it to log into Claude Code on that computer. His way: machines joined over Tailscale running T3 Code, so each time he says “use Claude Code” Grok Bot starts a T3 Code instance. For the setup itself his advice is “ask your agent.”
  • Why not Claude Code directly. (1) The phone UX: named bots instead of mixed-up threads, and he can call a bot and talk to it while driving (the calling feature was not available in Canada at recording). (2) Shared context: a fresh Claude Code session “has no idea” about his revenue or customers, while Grok Bot makes engineering decisions with that context. It once offered, unprompted, to reuse a Paper design system he had made for another family member.
  • Bots act across business and engineering. His VPC cloud project bot had another bot email a Canadian GPU provider asking for a discount, citing his YouTube audience, without being told to.
  • Security review. Use the best model you have, run /go, and ask for security checks. If Claude or Codex refuses because it thinks you are tricking it, open a new thread and reword. He also runs an “obliterated” model with guardrails removed in loops against his serious apps. He scales effort to the stakes: a free, login-less app gets less; anything with payments or sensitive data gets more.
  • Design before code (Paper MCP). Connect Paper’s MCP; screenshot designs from X or Mobbin and ask “can you build this on paper?”; have the bot build a design system (fonts, colours, type sizes, spacing, corner radius, button and card sizes) from a reference site, tweak it, and reuse it across projects. Leave comments on the Paper design, then tell the bot “address the comments”. Isenberg’s point: going straight to code is how you get the “AI smell”.
  • Content ideation from X. A YouTube bot runs at 9am and 5pm, scans X (he says Grok Bot’s X access “feels like unlimited”), and returns three topics with source tweets, title ideas and resources. He does not use AI to script (“people prefer people”). An editor bot turns his plain-English style notes into instructions he passes to his human editor.
  • Finance and advisor bots. A finance plugin connects bank accounts through Plaid. He runs business-finance, personal-finance, business-investor and personal-investor bots, plus an advisor bot that checks his “shiny object syndrome” against his goals and cash. On a rental property it flagged an “unrepresented buyers only” seller and a 17-year-old boiler and furnace. Other bots scan Canadian government contracts against his business.
  • Shopping. Daily eBay alerts for sealed-box GPUs. For a PS5 Pro, the bot emailed the seller, found it was a store two hours away, gave him the number and a call script, and completed the purchase once he gave it access. For shoes, it found options and emailed his wife for approval.
  • Give it the video. Ross says you can hand Grok Bot a video and it will pull the transcript and act on it.

Isenberg’s short-form note

Isenberg says he has been editing short-form videos with Opus 5.5: record a thought on his phone, give the recording to Claude Code with a framework for what makes a good edit, and post. He says some got “hundreds of thousands of views.” Ross’s version of the same idea: the expensive part now is the editing framework and domain knowledge, not the tooling.

Try It

  1. Audit your bots against his rule. One general bot for small ideas; a dedicated bot only for work you will keep shipping; delete the rest.
  2. Send your next plan to a bot with only “Review this” before asking it to build anything.
  3. Point a bot at Claude Code. Start with the easy path (log into Claude Code on the bot’s own computer) before networking your own machines.
  4. Schedule a twice-daily X scan for three topics with source tweets and title ideas, if your audience lives on X.
  5. Put a design system in a design tool before code. Generate it from a reference site, comment on it, then have the agent address the comments.
  6. Before connecting bank accounts through Plaid, decide what the bot may do with them. Ross gives his finance bots “full access” to his finances; Imoh, elsewhere in the wiki, withholds write access to his bank.

Open Questions

  • The subsidy figure. “12,000 worth of API token credits” for $200 a month is Ross’s estimate with no method given.
  • The “obliterated” model. He does not name it or show it, and says he cannot. Its use for defensive security review is unverified.
  • Per-bot computers. Ross says “every bot has a computer” and describes bots with “their own set of skills and its own computer”. He is a user, not SpaceX AI. See the open contradiction in Grok Bot.
  • Plaid finance plugin. Which plan includes it and what access scopes it requests are not stated.