Office Hours V: Paul's Compliance Scanner Meets Firecrawl, Painkillers vs Vitamins, What a Real Second Brain Costs, and Why Selling Got Harder

Paul's scanner needs to read a whole site - pixels before consent, ADA issues, licensing - and Claude had drafted a full architecture: React/Vite front end, a containerised Node + Playwright headless-browser worker on Render, a Supabase edge-function proxy, loosely coupled JSON so any check can be swapped. Dileep's second opinion: the crawling half is a solved product. Firecrawl.dev has search, scrape and crawl modes, its own MCP server and APIs, and returns markdown, a summary, a screenshot, the HTML and the links per page - he demoed it on apple.com. Free tier: 1,000 page credits a month; pricing scales for client volume and is 'the best I've seen'. Apify is the alternative MCP. Keep the swappable-checks design; point it at Firecrawl output instead of your own Playwright fleet. Lovable for the front end integrates with GitHub and Supabase - just change fonts, colours and copy when starting from a cloned design.

A painkiller relieves an acute problem - an elderly parent who cannot turn on Netflix - and needs little education to sell; a vitamin is nice-to-have and needs heavy positioning. Either way the message has to land on one of three human drivers: time, money, health. Cody's is time. Do basic research on the segment (where do elderly users and their families actually spend attention?) and match message, UX and channel to it. On the crashing APK: get the browser version fully working first; Replit can wrap it for Android later; distribution is not the hard part, positioning is.

In Karpathy's sense, a second brain watches your actual work and fills in the context you never typed; saving articles is a library, not a brain. screenpipe is the closest shipped example (captures your screen into searchable memory, SOPs, automations). The cost is the point: continuous observation is 'drinking tokens like a water fountain' - a 'token billionaire' pattern - so it only pays for people who want the AI to hold far more context than they currently give it. For ordinary session memory: Codex and Claude Code have shipped stronger built-in memory, and the Claude Mem skill (installed from its GitHub URL) manages it in both.

First, hosting: a product cannot depend on your laptop - move it to Hostinger. Second, the GitHub problem for non-technical builders: don't learn Git, drive it through a coding harness. Dileep demoed Codex connecting to GitHub with plain English ('help me connect to GitHub', 'check my GitHub repositories') and recommends Codex specifically for non-technical users. A dedicated Codex session was flagged to Niharika.

Writing-style skills (humanizer, the em-dash police) are use-case driven - personalise them and test whether output improves or degrades. Established engineering frameworks (Superpowers) should not be heavily edited. Method: have the AI extract the skill's salient points, then layer your own on top rather than rewriting wholesale. Related: a global personalisation setting asking for concise, limited-step answers is the cheapest way to cut Codex and Claude Code token use.

Dileep's diagnosis is the funnel, not the product: use Hotjar heatmaps to see scroll depth, above-the-fold engagement and bounce points; keep iterating copy; and post build-in-public content about the process, not the product. On pricing (Rs 6,000 launch / Rs 8,000 standard for a 12-week plan): free access went to relatives and friends, 'the wrong audience' - only strangers and cohort members give brutal feedback. Design a low-friction throwaway-price entry, then nurture up. And the line the hour is remembered for: everyone can build now, so distribution is the whole game. Vernacular voice agents, in passing: Sarvam, bolna.ai, smallest.ai.