2025-08-29
What if we ran a Claude Code agent 24/7 with full sudo permissions, and let it cook?
With LLM agents, execution should no longer be a bottleneck for shipping velocity. I want to build the fastest path from idea to market launch. Tell a chatbot your idea, test its prototypes, watch it deploy your prototype, profit.
I want to prove with this project:
AI Agents can build all of the business logic for most products. We see this already with people creating fast prototypes directly within ChatGPT or Claude, or paying for and using tools like Lovable and Base44.
However, there are two major pain points for agentic product iteration:
Run Claude Code on a VPS with full sudo permissions, connected to GitHub for human input. Send the VPS specs for new ideas and features, GitHub issues for fix requests, and general feedback for directionality. Focus on small and AI-context window manageable chunks, with less abstraction, to enable AI code generation tools to operate more effectively.
The short-term goal is to create a "junior engineer in a box" that, at full operational capacity, with a medium amount of human moderation, can build and ship one low-to-medium complexity digital product per week.
The "Shopify for building digital products".
Rather than generating apps from scratch, start from a curated library of proven app architectures (like Shopify themes):
We can take a three-layer system:
Layer 1: Business logic extraction
Layer 2: Template customization
Layer 3: Deployment and infrastructure
We don't need to generate thousands of lines of code for each variation of the same product. Instead, we should generate smart configuration files that drive pre-built, pre-optimized engines. WordPress/Shopify themes but for digital apps.
Assuming that the system works (of course), here's how this plays out:
The economics for digital products have fundamentally shifted. If it takes 20 engineers at $200k/year to build a single variation of a consumer/B2B digital product, the business needs to make more than $4M a year to break even on R&D costs (not even including ops costs). This makes it so that the business needs to target a large enough market to justify the cost.
However, we can now adjust the inputs in that formula - if it only costs you $20k/month in API calls and 1 engineer to do quality checks and code review, your breakeven point is now 10x lower than the incumbent. The economics have changed - you only need to find 3.7k people to pay you $10 a month to break even. If your system can build 10 products to ship, then you only need 370 people a month. You can play this game all the way to the bottom.
We have a single (or army of) junior engineer in a box that can grind 24/7 and churn out what I call "digital slop" (digital products built as variants of basic architecture, niched for specific audiences) that we can market and attempt to sell. Even in this scenario, we can build N products a year, spending human time on distribution. We become an AI slop shop, with our existential risk significantly lowered. We can get to a point where we're making software margins at a smaller volume, and then we play a scaling game.
We productionize our engineer-in-a-box and manage to automate enough of the workflow so that even the least technical person (think your friendly neighborhood aspirational MBA) can finally become the "ideas guy."
There's a hazy prosumer-style idea I have for monetization, where we position more as an aspirational product. We don't need to care if people can actually distribute and monetize the products they build with us, just that we can build functional enough products that people would pay for the prosumer product. Picks and shovels.
We're at an inflection point. When LLM tools become quality enough, software development (for non-super-sensitive products) becomes a race to the bottom. It becomes more a game of who can find distribution, lock down that niche market, and build up the walls of the walled garden first. Hyperpersonalization in products will make it so that tiny communities are economically viable target markets.
This isn't about whether software engineers are replaceable by AI - that's not the point. What matters is that the economics of software development have fundamentally changed. You can now execute and ship with significantly smaller teams, for significantly smaller target audiences. The barrier to entry has collapsed, and the first movers who can systematize this advantage will capture disproportionate value.
I'm going to do this regardless. For essentially under $50 a month I can spin up my own Claude Code agents in a Hetzner VPS and have them run 24/7. (I'm only spending $20 a month right now, but I'm adding some buffer in case other costs ultimately pop up.)
From my experience doing this so far, I know that this isn't an easy problem to solve at all - execution is king here and I am not worried whatsoever about people "stealing this idea." Do it! I want to see how other people try to solve this problem, and see what tuning, optimizations, and personalization come from other implementations.
My plan is to tune the full workflow and optimize it for my own build ideas, to have it handle the full product development journey. That process is so iterative and so personalized; it's already still so hard to make sure that the workflow is truly robust.
I think I can get a robust, automated workflow with a bare minimum UX built on top of it by the end of 2025, without help. I can ship 2-3 "shitty" products across a few different "templates" to prove that this can work.
I have one main ask that can be segmented a few ways. Help me speed up my iteration velocity!