Day 2: Custom GPTs and Gems, Markdown Prompting, Cloning Your Writing Style, MCP in Daily Use, a Vibe-Coded Calorie Counter, and the Consultant-to-SaaS Road

Use # / ## / ### to rank sections and *word* for emphasis, because the model reads formatting weight the way a human would. Five sections: Role (persona), Objective (the task), Context (why it matters - the MIT graduate who will not take a job seriously until told why), Instructions (numbered), Notes (everything else). Demoed on a customer-success agent for Outskill's inbox. It is the same framework BC9 Day 3 teaches for the same reason.

PROCEDURE: Create -> describe the bot in natural language (dictated) -> accept or change the proposed name and logo -> test in the preview -> refine in plain English (scoring, table output, ordering) -> save privately or publish to the GPT Store. Then the Configure tab shows the system prompt the wizard generated - the starting point for hand-editing into markdown form. Gemini Gems do the same for free.

PROCEDURE: (1) paste 9-10 of your own posts and ask for an extensive content-DNA report (hooks, body framework, CTA, signatures); (2) in a fresh chat, have an expert-prompt-engineer persona turn the report into a markdown Role/Objective/Context/Instructions/Notes prompt; (3) paste into a custom GPT or a Gem. Two tricks: tell the model 'Gemini scored 9.5/10 on this, you got 6/10, try again' and the second pass improves; when Gemini renders the markdown instead of showing it, ask for the output in a code block so the # and * survive copy-paste. Tested on AI bankruptcy, the four-day week and the H-1B fee.

PROCEDURE: transcribe the recording (Happy Scribe); write a prompt telling Gemini to answer learner questions strictly from the transcript, escalating to Perplexity/ChatGPT or a support email when it cannot; upload the transcript as Knowledge; test live. Then rebuild the same role / goal / instructions / knowledge in Lyzer AI, copy its Agent API code, and paste that into an Emergent prompt describing the chat UI ('black and white like Notion') - a hosted chatbot in 20-30 minutes. Finally VAPI: Assistants -> new -> model -> paste the system prompt -> attach the knowledge file -> pick a (cloned) voice -> Publish -> talk to it.

Level 1: Bolt AI as one chat client over OpenRouter's 500+ models and Ollama on his own GPU, with temperature, max tokens, top-p/top-k and penalties exposed. Level 2: MCP as the multiplier - Claude + Apify scrapes an Instagram profile's reels, ranks by engagement, pulls the top transcript and rewrites it as a LinkedIn post from one instruction; Goose (open source, on Claude) inspects a folder and recommends 50 GB of deletions without deleting; VAPI MCP places a real call to PV about pickleball. Level 3: a Flux LoRA trained on his photos plus a voice clone produce all of an Instagram account's content (200K -> 1.2M followers in eight months, '$100K+' sponsorship - his figures). Level 4: 'Jerry', the executive assistant that books meetings, drafts email (a live mismatch error shown), summarises a missed Fireflies meeting, and via computer use shortlists YC companies. Level 5: a CRM vibe-coded in three days for under $100 instead of a $120K HubSpot quote; a v0 ad-creative generator that opens pre-filled ChatGPT tabs. Estimated: L1 ~3 months, L2 ~4-5, L3 ~3-4, all five ~18 months at 10 hours a week.

Front end = what the audience sees; back end = everything unseen; API = assistant directors syncing front and back; database = the vault of raw footage; server = the team processing it; third-party integrations (shadcn/ui, the OpenAI SDK) = outsourced VFX - with the Henry Cavill moustache as what an integration mismatch looks like; auth = ticketing. Vocabulary given so non-coders can read what a builder is doing.

PROCEDURE: (1) ask Claude for a feature list; (2) research MyFitnessPal, Lose It, FatSecret and Cronometer in Perplexity and screenshot their premium features; (3) feed the screenshots back for a consolidated, prioritised feature table; (4) pick two or three core features - AI photo food logging and a calorie-goal calculator; (5) generate a PRD; (6) trim it - no database (local storage), no camera (web upload) - and compress into a guideline-style build prompt, not a prescriptive one; (7) paste into Replit (Genspark in parallel, which kept erroring); (8) debug: image analysis failing was the wrong OpenAI vision model, fixed by rolling back to GPT-4o; (9) add a text-input fallback via Plan mode; (10) revert to a checkpoint when an Apple-card restyle came out badly; Publish for a hosted URL. Rules stated: never interrupt a run, describe errors factually, one feature at a time - one 'small' feature touched ~20 files.

PROCEDURE as a path: Year 1 niche consulting, 2-3 clients, outcome-based pricing; Year 2 a small team, productised packages, a documented playbook; Year 3 delivery 80-90% automated with a 5-10% human exception layer on simple interfaces (Notion, Sheets, WhatsApp) - cost arbitrage against Accenture-class firms; Year 4 a SaaS with billing, dashboards and APIs (HubSpot's own path). The filter: people who live in one tool 6+ hours a day; deep-skill-but-patterned tasks; high-hourly-rate industries; specialised vocabulary; augment, don't replace; meet users where they are (WhatsApp, Excel, Google Forms, email). Then the 30-day playbook: ship one useful tool, document publicly, get testimonials, package an offer, build an audience, automate outreach.

Hunting grounds: Product Hunt 2017-2019 launches with 1,000+ upvotes that folded; Chrome Web Store extensions rated 4.5+ but unsupported; Indie Hackers projects plateaued around $5K a month on manual ops; pre-2020 YC demo-day companies with dead landing pages; GitHub repos with 1,000+ stars and no updates (check MIT vs non-commercial licences before monetising); subreddits with recurring complaints (freelance, legal advice, teachers).