Founder Stack Ep02 with Tanmay: Nothing Starts Where It Ends - Win the King, Start as a Service and Productise as You Go, Tech Is Not a Moat, Where Your Buyers Already Are, and Throw More Darts

You cannot plan year two on day one; stay open and adjust to what turns up. In 2023 Tanmay built an AI grading app in Bubble (teachers upload answer sheets and a rubric; the app grades) - 1,000+ users, nobody paid, and getting a thousand users is harder today because software supply has exploded. He had pitched five or six newsletters for a mention; only Ben's Bites ran it (unannounced - 50 visitors at once on his phone monitor). Months later he cited that mention when applying to Ben's Bites for an AI-educator role, got it, wrote 100+ tutorials and courses teaching businesses to use AI (teaching sharpened his own understanding), and a colleague's referral took him to a US agentic-AI startup as a forward-deployed engineer across medical and other clients - 'I was getting paid to learn'. The assets he accumulated were proof, skill, patterns and people; 'network is your net worth'. The principle: throw more darts; action produces information; set tight deadlines because time expands to fill them (Parkinson's law); 'a retard in motion beats a genius at rest' (his quote, as heard).

PrimeLoop began as 'whoever will pay, we'll build it' - broad, unsure of the market, small clients, unrelated problems, nothing to processise. The shift to one function - internal sales automation for the large sales teams of enterprises, which no agency was serving while outbound agencies crowded the external side - brought higher-ticket clients and credibility. Lesson one: pick a specialisation; most of the cohort already has one (the chat named marketing, finance, portfolio management, revenue ops, product ops, manufacturing, software, location services). Exercises: name the area where you know more than most; then your moat inside it (Caroline: aligning stakeholders around one project); then ask whether AI genuinely helps that moat - the answer may be no. Formula: Function x AI, delivered as a service (an 'AI-enabled HR agency' that builds internal matching tools) or a product (software for chartered accountants, his own field). Moats can be created: network, knowing the right things, data.

Credibility beats features: nobody reads your features page except the AI agents comparing products; people read the logos and the testimonials. Priority one, if starting over: bag the biggest names in the vertical, because selling downstream becomes easy. It is not as hard as founders assume - a Fortune-500 account executive signing up with a work email is enough. How: identify the kings of the market (mid-tech companies in Spain, say), follow the relevant people on LinkedIn and X, read what problems they post about, then write to them individually - never a template, 'you can't automate a human-to-human email'; offer a small free or paid trial without devaluing the premium positioning ('our services are premium; we'd like to give you a small trial because we think this is highly relevant'). Keep prices high: a $100 bottle is assumed better than a $20 one, and an enterprise buyer cannot justify a mediocre provider to their boss. Two viable shapes: twenty clients at $2k or four at $10k. Priority two: connect with people at adjacent and competing companies. Priority three: value flows both ways - networking without value does not stick. Agency pricing tiers: $1-2k a month is low tier and a headache, $2-4k mid, above $4k premium; start near $3k and climb; project pricing ($20-30k over months) when retainers do not fit.

Definitions: an AI product is self-servable (a link, a sign-up, the user does it); an AI service is anything where you are in the loop - automation work, consulting, your functional expertise delivered with AI. His take: start as a service even if the goal is a product. Four reasons: early cash flow (the North Star); direct customer contact - building alone in Claude Code or Lovable means guessing what the market wants; lower initial investment (no stack of subscriptions, no marketing spend); room to adapt - you learn the shape the product must take, and if the cash flow is not there you switch ideas instead of rebuilding, so you test more ideas per year. This is Paul Graham's 'do things that don't scale' - he rereads the essay monthly; Gumroad (The Minimalist Entrepreneur) and Amazon's early book deliveries ran manually first. 'Manual' means delivering the outcome with existing tools and your own judgment rather than shipping a system; then notice what repeats, build the internal system, and only then the self-serve product. Counter-cases he accepts: when the tech itself is the moat, or when cash flow already exists; and the reverse direction - a product company adds a service arm of forward-deployed engineers (Palantir; the startup he worked at), which charges more and keeps the market close. Q&A caveat from Gautam and the UK/GDPR side: when you learn from client work, write into the terms of service that the IP you build belongs to you; a lawyer before signing is money well spent, especially against exclusivity clauses.

When he taught the first Catalyst cohort he hedged that very hard tech might still be a moat; by C3 he does not: AI replicates products from a screenshot, so for 99.99% of businesses code cannot be the sole moat. Build something else: distribution first (people buy from people; even AI avatars trade on a human's trust), connections to the right people at the right companies second, credibility - the kings of the industry - third; also data others cannot get, industry knowledge that does not exist digitally, regional approvals or regulated access.

The fundamental question for a first customer: where do your buyers spend time online - and it is rarely the obvious place. Small-business owners on Facebook; tech buyers on X; some markets on Reddit or China's Rednote; tight communities like AIC Mastermind. Define the ICP, find the unexpected channel, add value before asking, and let the first happy customer start the snowball; demand was visible to him before he had any audience. PrimeLoop's channels in order: Upwork (an agency account; $90-100 in bids won a ~$1,500 first client, then two or three big logos; clients post what they will pay for - Harshit analysed ~8,000 AI jobs there; competitive now and some US clients only hire US freelancers, but still works); referrals ('shout from the roof'); cold-email outbound (booked meetings and clients, but the market was small so they did not scale it); content (long game, zero followers on day zero, start anyway). Do all channels; double down on what works. On X specifically: the most active AI ecosystem - follow the right people (start from his following list), scroll an hour a day and you are current; LinkedIn hears news three to seven days later (the Claude startup-programme story: a bot approved many of 100,000+ applicants, approvals were revoked within two days, X knew at once). Keep connections warm by sharing relevant updates; take initiative; have agency.

For knowledge work he keeps a folder (a project) per function - finance, hiring, internal operations, business work, executive decisions; thirty folders today. Principle: context stays close to the work, so an agent handling hiring never carries the whole company; this runs against the labs' push toward one agent with all context (Claude Projects, Dot), but it works. Benefits: tool-agnostic - open the folder in Claude Code, Codex or any new harness (add a claude.md or agents.md), so switching tools never means moving context; MCPs can be scoped per folder (Supabase only in the coding project), preventing cross-contamination; workflows and pieces are organised like code. Models and tools change; context compounds with the system.

Bonus from his AIC Mastermind talk: Hermes can be the main backend of an AI product (he built AutoCM on it) - customer requests, agent features, much that you would otherwise build from scratch; the maintainers are building it for that. Tools he uses daily: Factory Droid - a Claude-Code-like harness that is model-agnostic, so Claude, Codex and cheap open-source models all run in it (he uses Codex and Droid, not Claude Code); T3 Code and Conductor - free interfaces that run Claude Code and Codex side by side; OpenCode - an open-source harness with cheap plans and open models. Hosting: Oracle Cloud's always-free tier VPS (as heard: a few cores, ~12 GB RAM, 10 TB outbound) has run his Hermes for a month - with the caveat that Oracle can claw back free instances if eligibility lapses, so back the data up to paid storage or R2 and keep critical projects on a paid tier; he runs a second Hermes on an always-on local PC. The slides came from a front-end slides skill on GitHub (linked in the resources).