Ras Mic does not ask one giant chat to run his life. In this conversation with Greg Isenberg, he opens a set of Grok Bots organized around work that needs its own context: product engineering, design, video ideas, business decisions and shopping. Grok Bot is often the coordinator; Claude Code, Paper or a connected service does the specialist work. The useful lesson is how he decides what deserves a Bot, and when a human must still approve the result.
Idea or request → project Bot → specialist tool → reviewable result → human approval. Ras gives a durable project its own Bot, keeps a general Engineering Bot for small experiments and deletes temporary Bots when the job ends. For a first attempt, copy the separation of responsibilities, not his number of connected accounts or his spending.
Watch Greg Isenberg and Ras Mic
Host: Greg Isenberg. Guest: Ras Mic. The apps, revenue, pull requests, purchases and subscription-value figures below are their demonstrations or statements in the video, not independently audited outcomes.
One Bot per Serious Context, Not per Passing Thought
Ras's rule is context. Pluto and Bezalel are continuing products, so each has its own Bot. A screenshot extension he calls Still took two prompts and stays with his general Engineering Bot. IT and local-model work are standing roles; short-lived ideas can be deleted. He says Bezalel makes about $200 in monthly recurring revenue, but that is a self-reported figure, not proof of a repeatable business model.
The product architecture needs a correction. Ras describes a computer for each Bot, but Grok Bot's current documentation says Bots on one account share one persistent cloud computer: files, browser sessions and command-line credentials are available across them. Each Bot has its own screen and conversation. Those screens are work surfaces, not isolated security boundaries. Separate project folders and scoped app permissions still matter.
| Bot | Job in Ras's setup | Human checkpoint |
|---|---|---|
| Engineering | Turn an idea or reviewed plan into a repo and pull requests. | Read the diff, run tests and approve the merge or deployment. |
| Designer | Collect references, work in Paper and prepare a design system. | Approve originality, brand fit and layout before coding. |
| Content | Find three relevant topics and source posts twice a day. | Verify the primary source and choose what to publish. |
| Advisor | Put an idea against goals and, when connected, account data. | Check the underlying figures and make the decision yourself. |
| Shopping | Compare listings, contact a seller and prepare a purchase. | Confirm the seller, item, final total and payment. |
Engineering: Grok Bot Orchestrates, Claude Code Builds
For Pluto, Ras first asked his Bot to review a PDF plan. The Bot flagged decisions it would change; Ras supplied a non-negotiable requirement and pasted a friend's feedback. Only after he approved a revised plan did he ask Grok Bot to run Claude Code. He says the overnight work produced five pull requests that people were testing. That sequence is more instructive than the casual final command: review the plan, define constraints, delegate implementation, inspect the PRs.
He also demonstrates a Rust trading interface and an AI-governance prototype. The latter had been pushed to GitHub, but Ras explicitly said he had not even run it. A generated repo is therefore evidence of an experiment, not a working or secure product. His stated $8,000 to $12,000 of public-API-equivalent usage against a $200 subscription is his own rough comparison. It is not an API credit balance or a promise of future allowance; check Anthropic's current subscription and authentication terms before building a business around that economics.
His more elaborate route uses T3 Code and Tailscale across his own machines. Neither is required for a first Bot. More importantly, his security answer is to ask another model to inspect the work, sometimes an abliterated one. That can surface issues, but it is not a security certification. A free app with no login can still contain vulnerabilities. For any public release, review dependencies, secrets, network access, data handling and the deployed build; add specialist review where payments or sensitive data are involved.
Design: Make Decisions in Paper Before Writing Code
Ras sends references from X and Mobbin to his Designer Bot, then uses Paper's MCP connection to build editable layouts. He describes using that flow for his brother's clothing brand: choose a Christian Gothic direction, refine the design in Paper, then create a Shopify front end. He also uses Paper comments as a precise way to request revisions before an agent changes the site code.
A design system records the fonts, color choices, spacing, components and interaction rules an implementation should follow. It does not make borrowed artwork yours. Use a reference to explain a pattern or mood, then review the result for originality, accessibility and fit with the real audience. Paper's official guide explains how agents can read and write a design file through MCP; connecting it still grants access to that file, so confirm the correct project is open before asking an agent to edit.
Content: Three Ideas at 9 a.m. and 5 p.m.
Ras's YouTube Bot checks X twice daily and sends three potential video topics with source posts, title ideas and further resources. He does not publish all three, and he usually scripts his own videos. His Editor Bot, meanwhile, helps turn his informal visual feedback into a brief for his human editor; at recording time he says no video made with that Bot had gone live. This is a research and briefing workflow, not evidence that his channel is fully autonomous.
Grok Bot's routine guide recommends proving a task once before scheduling it. A useful scheduled brief should name the audience, time zone, primary sources, output format, what to do when no reliable news appears, and a rule against auto-publishing. Public X search is available without connecting an account, according to the official app guide. A viral post still needs verification before it becomes a factual video.
Business and Finance: Context Without Surrendering Judgment
Ras's Advisor Bot pushes back when he wants to chase a new venture. For a real-estate listing, it argued that the purchase would compete with his other goals and flagged questions about representation and older heating equipment. He says his finance Bots have account connections through Plaid, letting the Advisor consider transaction context. This is his personal workflow, not an independently checked investment analysis or a recommendation about that property.
Bank connections are a different risk class from reading public posts. Plaid explains that connected apps receive data you agree to share. Check the exact connector, account scope, retention and disconnect path before authorizing it. Keep the Bot's role to organizing records and posing questions; reconcile material numbers against statements and use a qualified professional for consequential financial, tax or legal decisions.
Shopping: Let the Bot Search, Keep the Checkout Decision
Ras has a Bot look for sealed GPU listings. He says it contacted a seller about a PS5 Pro, found the store in Ontario, supplied a phone number and bought the console after he confirmed. For another purchase, it sent shoe options to his wife for feedback. The transferable pattern is search → verify → approve, not unattended spending.
There is an important product detail here: Grok Bot's Vault documentation says a saved card can be used at checkout without asking on every use. If you require approval for every payment, do not save a card in Vault; use a one-time secure form and inspect the final order. Auto Review can pause consequential actions, but its approval applies to the proposed action, not earlier work the Bot already completed. Set an explicit stop before contacting a seller or checking out.
A Smaller Setup You Can Copy Today
- Pick one low-risk job. A sourced content brief is easier to verify than a finance or shopping Bot. Use one Bot, not five.
- Give it a role and a stop rule. Name the audience, sources, output and forbidden actions in the first message. The official setup guide recommends one job per Bot.
- Run a one-time test. Open every source link, check the date and ask it to explain any uncertainty. Revise the instructions until the output is consistently useful.
- Save the method, then schedule it. Turn the proven task into a skill or routine only after you can describe the success check, failure behavior, time zone and likely usage.
- Add specialist tools later. Connect Paper only for a design project, Claude Code only for an authorized repo, and account data only when a concrete task justifies its sensitivity.
Starter brief, adapted from the episode: “You are my content-research Bot for [audience]. Once, find up to three developments from the last 48 hours using original announcements where possible. For each, provide the source URL, publication date, what is confirmed, what is only a claim, and one possible video angle. If you cannot verify a claim, say so. Do not post, email, buy, create a routine or connect an account. Ask me which topic is worth pursuing.”
After that test, you can ask for a twice-daily routine like Ras's. Keep its first runs supervised. The point is to let the Bot prepare options while you retain the editorial, technical and financial decisions.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Intro and the five areas |
| 01:55 | Engineering Bot and project contexts |
| 07:57 | Why Grok Bot orchestrates |
| 14:01 | Grok Bot with Claude Code |
| 19:43 | Security checks |
| 21:54 | Designer Bot and Paper |
| 31:12 | Content Bot and video ideas |
| 37:19 | Business and Finance Bot |
| 42:41 | Shopping Bot |
| 44:53 | Closing thoughts |
Sources and Limits
Checked 11 November 2026. The episode and supplied transcript establish what Greg and Ras said and showed. Current behavior and boundaries were checked against Grok Bot's computer guide, routines guide, approval guide and Vault guide; Paper's MCP docs, Claude Code terms and Plaid's connection explainer inform the caveats. I did not inspect Ras's private Bots, repositories, bank data, purchase records or billing. The starter brief and checkpoints are JQ AI SYSTEMS editorial guidance, not quotes from Ras.