Office Hours I: Vetting an Idea, Niching Against the Frontier Labs, and Paul's Aircraft-Parts App (PWA, and 'What If I Get Hit by a Bus')

Harshit's three-part check, in order. Run a deep-research query on Gemini or Perplexity describing the product and asking what exists. Then read what people say about the nearest tools on Reddit and X - the complaints are where your differentiation lives. Then search Product Hunt for launches in the last months and read the comments under them. If all three come back empty, the idea is either new or nobody wants it, and the next step is a quick landing page to find out which.

Gautam's strategic question gets the longest answer of the hour. The labs will own high-security and regulated work, and any generic AI-wrapper product dies with the next model release. What they cannot serve economically is the small and mid-sized business - their engineers cost $250-300 an hour and sit in Western time zones. So: niche down to one vertical you already understand (the host's example is automotive), go deep on that industry's actual problems, and never build a product whose whole value is a thin layer over a model.

Paul described a repair company for aircraft parts that wants to photograph incoming parts and again after repair, from a tablet or phone, with the images going straight to their own server. The answer: build a responsive web app and ship it as a Progressive Web App. It installs to the home screen, uses the camera, works on both tablets and phones, and never goes through Apple or Google review. The transmission question ('ways to transmit it across the network that makes sense to a server') is ordinary - the PWA posts to the server's API.

Harshit does not pretend to a clean answer. Do not hand over the codebase unless the client pays for it - it is your leverage and your product. Train their people on operating (not building) it. Or build automated monitoring and alerts so the system tells someone when it breaks, standing in for your personal availability. Paul pushed: 'how do I get a backup colleague that I can trust?' - and the honest reply is that a client holding the application can always find 'a person like yourself', which is a comfort, not a plan. 'We'll talk more about it later.'

The $20 Claude plan is 'useless' for real coding because its session limits end the work mid-task. Harshit's options: the $100 Claude plan, or Codex, which he rates at comparable quality to Opus (heard as Opus 4.7 vs GPT-5.5) with higher rate limits. Windsurf and the other editors are just paid wrappers over the same models. Separately: CAD-grade architectural drawings from image models are not yet possible - that needs a specialist engineer, so post in the cohort group.

The participant's app was an exported HTML file talking to Firebase and Grok. Harshit's verdict: distributing app logic (and any keys) as a local file is 'not safe' beyond personal use - host it, or if the point is the output, ship a PDF instead. For guardrails generally: read the labs' published guidance, search Perplexity for current practice, and wait for the dedicated security session. Swaminathan's climate-product security fears get the same deferral plus a rule: a non-technical founder needs an engineer on call for security response, and even Anthropic engineers make mistakes.

Ask your own coding agent to write an architecture and guideline document, then redact it to what operations and support actually need. Stand up two agents - one for ops, one for support - trained only on the redacted material, and attach each to its own inbox (a Hermes-style email agent was the example). The team works through the agents; the backend stays yours.