Office Hours XIV: Story Before Landing Page, Scoping an Agentic-Web Pitch, the Cost of Compute, Automating Past Google, and the Idea Bank

Jane, an author new to online work, built a testimonial site to meet 'people who resonate with the ideas of healing, feeling, art, creativity'. Harshit: 'I see a lot of text' - visitors skim, so each scroll position needs one element in focus. Asked what the agenda is, she says she is 'figuring it out as I go along'. His advice: 'forget about the landing page as of now, just pull up a Google doc', dictate or ramble what you want to teach and which pains you solve, refine it, and only then turn the page into a storyboard where a visitor scrolling top to bottom connects with her story and what she can do for them. Her video belongs in the hero. 'It's not about design or elements or using AI the right way; it's actually about telling the right story.' For a writer the page is secondary: publish weekly (Substack), end every post with a call to action, and you 'increase your luck surface'. He points her to writers whose own sites are worth studying (Dan Koe, Kapil Gupta). By the end of the hour another member had built her a scrollytelling site around the video.

After Cody's demo Harshit says not to worry about 'mcp a2a webmcp whatever'; the question is 'what's my vision and what is the quickest way to drop this product into my target audience'. Serving every non-technical person across 'banking travel health government' is something 'even companies like OpenAI won't be able to do' - scope to one action, since 'even booking tickets is a big market'. The goal is 'these five people to use my product', then feedback, because 'nobody really knows how everything is going to pan out' and stacking features rests on assumptions. Consumer personal-assistant products 'can crush you anytime'; B2B is the better space, where buyers 'do not want to lock in on any particular model provider' and ask for ROI and 'value from day one'. Show 'one single linear story' - several flows at once confuses people. Cody's one-liner ('enable a giant amount of masses') gets the reply: show them how; he plans 20-30-second videos and concedes hackathon complexity does not survive the real world.

Cody's friend frames the idea as intent-based computing: most people, 'my mother', cannot drive booking.com through an agent, so one voice agent that remembers a person's preferences should search, show options and buy on a yes. Cody says the site is 'literally connected to what you taught us' - the Sunday WebMCP session - and loads A2A, WebMCP and a third protocol Whisper renders as 'UDP' (unresolved). A live view shows the packets as a research paper is processed, and a dashboard compares steps and tokens with and without it. His thesis: 'there should be no more browsers' for agents, so a site shows humans one face and bots another (a 'CLI mode'). 'Recipes' are skill-like flows that call tools in a fixed five-to-six-step order; a recipe store and a mobile app are planned. He states there are no beta users and the dashboards are demo data. A side claim: agent traffic breaks web analytics (server location, scraping, inflated impressions), so a partner is building agent-aware analytics - Harshit: 'that's a good problem statement'. The finished intent demo (bank balance, add beneficiary, 'traditional vs Elastic Web Concept') drew: 'I just need to see a real world use case.'

He argues owning compute 'is not eco-friendly... sustainable', regrets not buying a Mac Studio ('four thousand dollars up'), and notes Hostinger rents GPUs by the hour and deletes your space when the hours run out. Harshit agrees it is 'very hard' for individuals while enterprises 'strike a deal', and expects the convergence 'at the lab level'. The cohort is 'the top maybe 0.01 percent' of users: consumers are fine on $10-20 plans, entrepreneurs end up on $100 subscriptions and API costs climb fast. He predicts two or three main players within three to four years. Pooling credits across the cohort runs into labs being 'very very strict about their licensing and partnership deals', and self-hosting never matches their API price. The member's proof of the problem: 'I burned 4,000 credits in less than 15 minutes' of a 65,000-credit gift. Harshit: 'that's why we have the mastermind sessions.'

A member notes that everything the cohort learned in n8n lives inside Google: Yahoo auto-forwards only on a paid plan, Meta's 24-hour window blocks automated messages, and WhatsApp and LinkedIn block accounts. Harshit's first move with a client: ask which tools they use and whether there is an API, find a trustworthy MCP, and 'build some form of skill-based mechanism' - which 'makes sense only at enterprise level not at individual level'. On WhatsApp: 'they block me', and the business licence takes 'a week, 10 days', so take clients to Telegram and automate there - 'much much safer and easier'. There is no API from Telegram into WhatsApp; the 'hackiest way' is both open in a browser with Claude Code or Codex using browser-use to sync them. The member already drives WhatsApp through Grokbot with computer-use on his own machine. Harshit warns platforms can block such agents, as Amazon has blocked a rival assistant.

A member runs Hermes on an Ollama Cloud plan (models heard as Kimi K2.6 and GLM 5.3 Flash) to build wealth apps and sites, and auto-restarts with a handoff around 64,000 tokens of context, yet still loses instructions and table detail - for example repeated scraping rules across ten sites. Harshit: context 'is something that you will have to keep clearing'; Codex and Claude Code compact well, and a skill can compact at about 80 percent of the window or open a new thread at about 60 percent. But the structural fix is decomposition: because you know your daily tasks, create skills that break them down, give each website to its own sub-agent in its own context window, and return 'just the answers and insights' to a thin orchestrator chat that stays high level - the way software teams work, since an 'entire code base can't be moved into one context window'.

Jane says learning AI has been 'so enjoyable' but so many possibilities have opened that she cannot keep track. A member answers that it is 'overwhelming, definitely', missing a day feels like missing a lot, and two or three active communities help - including AIC Mastermind, 'Prafulbhai's creation' - because you at least learn what is happening; 'you cannot be on top of everything'. Jane: 'no one talks about this', and she felt it 'so intensely for the first half of Catalyst'. Paul Berg adds that everyone goes through it. His method is an 'idea bank': 'I just had Claude create something where I can put ideas into it' - ideas are out of his head, remembered, fleshed out later, and acted on one at a time. His warning: the work 'can become a mission in and of itself... almost addicting' ('what if I do that?'), so stop and ask 'what am I really trying to accomplish? What is my main goal here?' Echoing Harshit's earlier advice to spend time on the writing and the audience rather than the landing page: 'the mission is not necessarily the media... your end game'. Jane's site, built around her story and video, arrives a minute later.