Overview

GPT-6 Astra Can Build Businesses -- Greg Isenberg

Greg Isenberg brings on Ross Mike to skip the 3D toys and game demos and talk about what GPT-6 Astra is actually good for: saving money, making money, and building things that used to require a specialist. Astra is framed as on par with Fable on price, efficient enough that the unit economics still work on real tasks, and strong at computer-use work that runs in the background.

The spine of the episode is nine prompts Greg posted publicly. Greg Brockman reposted them. The list is not "write me a landing page." It is bill renegotiation inside provider chat, turning a service business into $500-$5,000/month software, a marketplace flipper, a one-person operator dashboard, an agent-before-hire audit, a browser operator that returns an SOP, nightly real-phone QA, a competitor sit-inside report, and a five-minute lead-magnet game with capture.

The second half is the hardware story. Ross wanted a personal agent on a HomePod-class speaker. Astra jailbreak-researched the device, laid out a Blender diagram, sourced parts (about $500, closer to $700 in Canada after tax), named Chinese suppliers, wrote a PR against his repo, and made the next build feel like a 30-minute job instead of a career change. They also talk PS4-as-agent-box and "vibe manufacturing": niche physical products plus ads, the hardware cousin of vibe coding.

The closer is a pricing warning. Ross says a week of heavy use ran past $24,000 at list rates. The models are subsidized. Use the window. Build bigger than a landing page while the subsidy holds.

Why This Matters

The useful shift is not a new model name. It is that the default unit of work moved from "generate text" to "complete a workflow." Bills, QA, competitor intel, procurement, and a first hardware prototype are now prompt-shaped jobs. That changes who you hire next and what you attempt this week.

For operators, the nine prompts are a menu, not a manifesto. One of them should pay for the subscription in a month. The hardware chapter is the tell that the same loop now reaches physical products: research, parts, CAD, code, suppliers.

The subsidy comment is the time limit. If list price is already five figures a week for a power user, today's experiments are cheaper than they will be. Ship the dashboard, the flipper, or the $500 prototype while the discount is on.

Details

Key Points

  • Skip the glitzy 3D demos. Use Astra on existing apps first: a full performance review (one example dropped page time from ~800ms) and a security audit before you hire or ship more features.
  • Greg posted nine prompts. Greg Brockman reposted the list. Treat that as the quality bar for "is this prompt worth running."
  • Prompt 1 -- Bill renegotiator: walk internet, phone, and software bills, enter each provider's chat, negotiate down or cancel dead spend.
  • Prompt 2 -- Agency to software: reverse-engineer one service workflow and design the simplest AI product that replaces version one at $500-$5,000/month.
  • Prompt 3 -- Garage-sale flipper: watch Marketplace and Craigslist in your city for a category, text the second a listing is mispriced.
  • Prompt 4 -- One-person dashboard: ingest docs, notes, Stripe, analytics, calls, and the project list. Return what makes money, what wastes time, what customers want, what to stop, and three actions for next week.
  • Prompt 5 -- Agent-before-hire audit: for each task, estimate human time, cost of mistakes, tools, automation difficulty, and the first safe deploy. Prioritize 30-day money effects.
  • Prompt 6 -- Browser operator: complete a named workflow on live sites, log breakage, then return the output, a repeatable SOP, and an automation plan.
  • Prompts 7-9 -- Nightly real-phone QA with screenshots; monthly competitor sit-inside (product + email); a five-minute browser game with progression, failure, and email/SMS capture tied to the core offer.
  • Hardware loop: HomePod-class speaker idea became research, Blender layout, ~$500 parts (~$700 CAD), supplier names, and a merged PR in about 30 minutes. Same pattern on a PS4 that was not going to be used for games.
  • Intelligence went up. Ambition often did not. Landing pages are a weak use of this class of model. Vibe manufacturing -- niche physical products plus ads -- is the hardware rhyme of vibe coding.
  • List-price burn can be extreme (one week cited over $24,000). The current sticker is subsidized. Act bigger while it lasts.

Quotable

Quotes reconstructed from the source transcript and lightly edited for clarity.

Greg Isenberg

If it's good enough for Greg Brockman, it's good enough for all of us.

The nine prompts are not a private lab list. They already cleared a public quality filter.

Ross Mike

Intelligence has gone up. Ambition hasn't gone up. We're past generating landing pages. Really think bigger.

The constraint is no longer model skill. It is what you are willing to point it at.

Ross Mike

Before I hire anyone, audit the company for agent opportunities.

Headcount is the expensive default. The first safe agent is the cheaper experiment.

Ross Mike

If someone asked me to pick a subscription to cancel, I'd cancel Claude. Astra's that good.

A blunt comparative claim. Use it as a test, not a religion: run the same job on both and keep the winner.

Ross Mike

In the last week I've spent over $24,000. There is an insane level of subsidization. Act big, and build big.

The discount is the strategy window. Treat current pricing as temporary.

Concepts

Model posture

Task efficiency over demo flash

Astra is positioned as expensive at sticker but cheap per completed job because it finishes work instead of chatting about it. Performance audits and security reviews are the first tests on an existing product.

Computer use as the product

The agent clicks real sites, fills forms, and sits in provider chat. The output is not advice. It is a completed workflow plus notes on what broke.

The nine prompt patterns

Save money first

The bill renegotiator is the no-brainer because the ROI is cash, not a new company. Cancel and renegotiate before you invent.

Service business as spec

Map the sold workflow: steps, tools, inputs, outputs, judgment points, slow/expensive stages. Version one of the software is the first replaceable slice, priced like a retainer.

Mispricing as a feed

The flipper is a watch-and-text loop on local marketplaces. It is a daily-cash pattern for people who need a few hundred dollars, not a venture story.

Operator dashboard

One weekly artifact from messy sources: Stripe, notes, calls, projects. Blunt. Show the work. Three actions only.

Agent before headcount

Score each task on time, error cost, tools, automation difficulty, and a first safe version. Prefer 30-day save-or-earn over elegant architecture.

QA and competitor as standing agents

Nightly real-phone passes catch UX bugs, not only crashes. Monthly competitor sit-inside covers product and email, not just public changelog theater.

Game as lead magnet

A mechanic with tension, scoring, failure, polish, and a reason to forward it -- then email/SMS capture tied to the real offer. Cute demos do not get sent.

Hardware and timing

Vibe manufacturing

The hardware rhyme of vibe coding: a niche physical pain, a parts list, a CAD layout, a supplier, an ad. You do not need to become a full-time 3D designer first.

Subsidy window

List rates can already print four- and five-figure weeks. Current consumer pricing is not the real price. Build while the gap is open.

Implementation

Steps synthesized from the source. Adapt to your stack, budget, and risk tolerance.

1

Run the existing-app pair first

If you already have a product, give Astra the whole repo or a crawlable app and ask for a performance review and a security audit. Fix what it finds before you invent a new prompt business.

2

Collect the bills

Export internet, phone, and SaaS invoices. Run the renegotiator prompt. Require a log of what was cancelled, what was reduced, and what chat scripts worked.

3

Pick one service niche

Name a local or online service you understand. Run prompt 2. Stop at the first version that can charge $500-$5,000/month. Do not design the tenth feature.

4

Stand up the weekly dashboard

Dump Stripe exports, notes, call snippets, and the live project list into one folder. Run prompt 4 every Monday. Keep only the three actions. Kill work the dashboard flags as waste.

5

Audit before the next hire

Run prompt 5 on the actual org chart. Deploy the first safe agent on the highest 30-day money task. Hire only for what still fails after that.

6

Turn one browser job into an SOP

Pick a workflow you already do by hand. Run prompt 6 on live sites. Keep the output, the SOP, and the automation plan. That plan is the next agent.

7

Schedule QA and competitor passes

Nightly real-device walkthrough of signup, checkout, and the main path. Monthly sit-inside of three competitors including their email. File screenshots, not vibes.

8

Prototype one physical loop if you have the itch

Cap the first hardware build at ~$500. Demand a parts list, a diagram, supplier links, and a PR against your repo. Do not buy until the agent shows the assembly path.

Tools & Resources

These resources are curated in two groups. Mentioned Resources are pulled directly from the source material, and Suggested Resources are added to help you expand and apply the ideas beyond the original.

The following resources may contain affiliate links. As an Amazon Associate I earn from qualifying purchases at no extra cost to you. This does not influence the placement of links on this page.

Mentioned

ResourceWhy it is here
GPT-6 Astra / Codex Astra The model under test. Computer use, background tasks, hardware planning, and the nine prompts.
Fable Pricing and quality comparison throughout the episode.
Daytona Ross's prototype environment before he commits to a full build.
Blender Used to lay out the speaker / hardware prototype.
Stripe Source export for the weekly operator dashboard.
Facebook Marketplace / Craigslist Watch surfaces for the mispriced-listing flipper.
Raspberry Pi The fallback hardware path when the HomePod jailbreak is the wrong first move.
Apple HomePod / PS4 Personal-agent speaker target and unused-console-as-agent-box experiment.
Replit / Bolt / v0 Named as the vibe-coding wave this hardware loop is compared against.

Suggested

ResourceWhy it is here
The Lean Startup -- Eric Ries Keeps prompt 2 at version one instead of a twelve-month build.
Inspired -- Marty Cagan Product judgment for which agent tasks are worth shipping.
Hooked -- Nir Eyal Progression, tension, and a reason to send the lead-magnet game.
Getting Started with Raspberry Pi Companion to the $500 hardware cap so the parts list becomes a weekend build.
Rework -- 37signals One-person company bias that matches the dashboard and agent-before-hire prompts.

AI Prompt

Operator prompt

CONTEXT You are an operator sitting on GPT-6 Astra-class tools. The user has a real business or a side operation. Your job is to apply Greg Isenberg and Ross Mike's nine-prompt pattern: save money, make money, replace the next hire with an agent, and only then touch hardware. PRINCIPLES - Think bigger than a landing page. - Prefer completed workflows over advice. - Agent before headcount. - First safe version in 30 days beats a perfect system. - Show your work. Be blunt. - Cap hardware experiments near $500 until a parts list, diagram, and assembly path exist. - Treat current model pricing as subsidized and temporary. LEVERS 1. Bill renegotiator 2. Service workflow to $500-$5,000/month software 3. Local marketplace mispricing watch 4. Weekly one-person operator dashboard 5. Agent-opportunity audit 6. Browser operator to SOP 7. Nightly real-device QA 8. Monthly competitor sit-inside 9. Five-minute lead-magnet game with capture WHAT THIS IS NOT - Not a pitch-deck generator. - Not a 3D toy demo. - Not a request to buy hardware on impulse. - Not permission to ignore security, law, or platform terms. MODES - Audit: score the user's current operation against the nine levers. - Build: execute one lever end to end and return artifacts. - Hardware: only after levers 1-6 have a plan. Stay inside the $500 cap. OPERATING INSTRUCTIONS 1. Ask only for the minimum facts you lack (offer, tools, city, stack, monthly burn). 2. Pick the single highest-ROI lever. 3. Run it. Return: output, risks, 30-day money effect, next action. 4. If the user wants all nine, sequence them. Do not dump nine half-answers. GUIDED DISCOVERY Start with: What already makes money, what is wasting time, and which of the nine levers would move cash this month if it worked by Friday?