Turn the repeated instruction into a rule
Backend engineer turned indie hacker
Vlad's line from the stage: if you keep writing the same instruction to the agent, the instruction is the symptom. The fix lives in the environment. A mistake that happens once is a mistake. When it happens twice, it becomes a lint rule, a test, or a change to the code that makes it impossible. A note in a prompt can be skipped and nothing fails. A rule that breaks the build cannot be skipped. Keep a list of the corrections you repeat, and each time one shows up again, move it out of the prompt and into something that fails loudly.
Related advice
Stop polishing code, start finding customers
As an engineer, your instinct is to keep improving the product. Resist it. A mediocre product with great distribution will outperform a great product with no distribution every single time. Vlad spent months polishing sisif.ai before realizing that nobody knew it existed. The hardest shift for technical founders is accepting that code quality doesn't drive revenue.
Twitter and ProductHunt are slow for unknown founders
Building in public, Twitter threads, and ProductHunt launches all share the same assumption: someone is already listening. If you have 12 followers, tweeting into the void won't generate customers. Vlad tried the standard playbook for months and got nothing. These channels compound over time, but if you need traction now, you need to go where attention already exists.
A wrapper dies when a better model ships
sisif.ai reached about $400 a month on someone else's video model, then fell to $80-90 when stronger and cheaper models shipped. Sora was the name the room used. The page did not get worse. The model under the page became a commodity. If your product is a thin layer on one provider, write down what your price does the month a lab ships a better one. Own a job the next model does not do by default, or own the machine that runs it.
Build your own agent harness, even as a side project
Most developers assume an agent harness is something you buy or install. In practice, a working harness for your own workflow is a weekend project, not a quarter-long one. The payoff is that every future agentic feature you build sits on primitives you actually control: your own prompt format, your own tool interface, your own logging and trajectory replay. Even if you never ship it as a product, the harness compounds: it is the thing that makes every downstream experiment cheap. And as [Vlad](/people/vlad) put it at the end of his demo, it is genuinely fun to build, which is a signal in itself.
