Your AI Lunch Guy

April 2026 21 min read

Building Your AI Lunch Guy in 5 days

I built a whole business in 5 days, solo, without writing a line of code by hand. By the end it had a live website with booking, payments, automated intake, AI session prep and a full post-session workflow. This is the day-by-day of how it came together.

๐Ÿ’ฌ The post that kicked it off โ†’ Building in public: 5-day AI business challenge

๐Ÿฅช The final result โ†’ yourailunchguy.com

The premise

I'd been helping friends get set up with AI agents over lunch, literally sitting next to them at a cafe, installing tools on their laptop and building their first workflow together. One buddy got a ClickUp-to-Wise automation over fish tacos in Hayes Valley, and another had his inbox sorting itself in 45 minutes.

So I decided to turn it into a real thing: an in-person service where people book a lunch in San Francisco and get their first AI agent workflow built with them, hands on. It was also a 5-day experiment in building in public. The challenge was to build the entire business with AI tools (brand, website, booking flow, payments and automation), solo, while juggling a toddler.

Your AI Lunch Guy screenshot

The main tools were Claude Code (Opus 4.6), Notion AI, VS Code, Google AI Studio and Wispr Flow, plus a bunch of integrations I'll get to below.

Day 1: Strategy & planning

Tools: Notion AI (Claude Opus 4.6), Wispr Flow

Spending the whole first day (about 5 hours) on strategy might sound slow for a 5-day build. But the stuff that slows you down on days 3, 4, 5 is almost always something you skipped on day 1.

If the brand voice isn't locked in, you end up rewriting every page of copy. If the target audience is unclear, your messaging talks to everyone and lands with nobody. Every shortcut on day 1 turns into a tax you pay on every day after it.

Picking the name

I started with "Your AI Lunch Guy" as a working name and pressure-tested it. I gave Notion AI the full context (service description, vibe, positioning, target customer) and asked for an honest evaluation plus 15-20 alternatives. We went through 50+ options across 7 rounds of literal names, wordplay names, abstract names and personal brand names.

"Agents with Race" sounded premium, but "agents" still means nothing to most non-technical people. "Prompt & Plate" was clever but could read as a food brand. "Bots & Burritos" was memorable (nobody forgets that name) but too niche.

I landed back on the original. "Your AI Lunch Guy" won because it sounds like a nickname you'd give a friend. "Yeah I have an AI Lunch Guy, you don't?" That casual "I have a guy for that" feeling is exactly the vibe I wanted. 45 minutes total.

Voice and tone

Everything compounds off the tone decisions you make early, so I gave this about an hour. That felt like a lot for "just voice and tone", but every piece of copy after it went faster because the guardrails were already set.

I defined it across 4 axes (90% casual, 70% playful, 80% accessible, 75% conversational) and then 6 core messaging principles. The big ones were show don't pitch, keep the stakes low, talk like a friend but prove like an operator, and lean into lunch because lunch is the differentiator.

I also built a full "do use / don't use" language guide. Words like "sit down with", "set up" and "build together" are in. AI hype language ("supercharge", "unlock" and friends) and corporate jargon are permanently out.

Who it's for

My first instinct was to lead with "founders and operators". AI pushed back, and I agreed. Leading with that narrows the audience and scares away the people who'd get the most out of it: curious, non-technical professionals who use ChatGPT but haven't touched agents yet.

We defined 5 segments: the AI-curious professional, the curious founder, the ops operator, the team buyer and the "my friend told me about you" person. For each one we mapped who they are, how they use AI today, their pain point, what success looks like after one lunch and how they'd find me.

Writing the website copy

With the voice locked and the audience defined, the rest moved fast. I wrote each section in a separate conversation and fed it the brand voice doc plus all the previous copy so everything stayed consistent.

The hero went through 4 versions before landing on "You figured out ChatGPT. Now what?" That headline works because it meets the reader where they are. Most people have figured out ChatGPT, but "agents" still sounds vague and a little intimidating, and the headline names that gap without making anyone feel dumb.

I also wrote the about section, how it works (4 steps), use cases (4 audience segments), a comparison table ("Why not just YouTube it?"), the FAQ, testimonials (placeholders for now), a bento grid ("What you'll walk away with") and the CTA.

Pricing, booking flow and intake form

Pricing is a $5 reservation fee (it filters out no-shows) plus $200/hr billed after the lunch.

I mapped the full 11-step customer journey from discovery to post-session follow-up. One fix I caught along the way was that the original flow sent the invoice before the follow-up email, so I swapped them and the customer gets the helpful resource doc first and the invoice second.

Then I drafted 9 intake form questions in the brand voice, like "So what do you do all day?" instead of "What is your job title?" The money question is "What's the thing you do every week that makes you go ๐Ÿ˜ฉ?" because it surfaces the exact pain point I'll build around in the session.

Design direction

I used Wispr Flow's website as the inspiration. Warm instead of cold, with cream and off-white backgrounds in place of sterile white, 2 fonts (serif for headings, sans-serif for body), soft purple for buttons, deep green for contrast and a pop of orange used sparingly. No logo, and rounded corners on everything.

End of Day 1: Name locked, voice defined, audience mapped, full website copy written, pricing set, booking flow designed, intake form drafted and design direction chosen.

Day 2: Website build

Tools: Claude Code (Opus 4.6), VS Code, Playwright MCP, Notion MCP, GitHub, Wispr Flow

The whole point of a full day on strategy was that when I opened Claude Code this morning, I was just executing. Every piece of copy, every section and every design direction was already documented in Notion, so Claude Code only needed to read it and build.

Setting up the dev environment

I created a new project folder, opened it in VS Code, enabled the Claude Code extension, linked Git to a GitHub repo and turned on 2 MCPs (MCP servers are plug-ins that let Claude Code reach other tools):

  • Notion MCP: lets Claude Code read my Notion pages directly. All the copy, brand voice and design specs from Day 1 were available to it without me copy-pasting anything.
  • Playwright MCP: lets Claude Code drive a browser. It can open the site, take screenshots and actually see what it built.

15 minutes total, and most of that was making sure everything was talking to each other.

Wireframe first, design later

This was the biggest decision of the day. I used Wispr Flow to talk my first prompt into Claude Code: build the scaffolding for this website, look at this Notion page (my full copy page), and ask clarifying questions before starting. No design yet, just a wireframe.

I turned on plan mode in Claude Code so it came back with a plan before writing any code. I reviewed the whole plan carefully, went back and forth, and only accepted it once every section was right. Fixing things in plan mode is way easier than fixing them after they're built.

Within 45 minutes of that first prompt I had a working wireframe of the entire site running on localhost, with every section in real copy and full navigation. Ugly as hell, but structurally complete.

Then came about 90 minutes of rapid-fire refinements through Wispr Flow: center the banner, make the hero text much bigger, show one tile at a time with pill buttons, stack the How It Works steps vertically and stagger them, make the before/after cards overlap. The loop was simple. I'd look at the site, spot something off, say it into Wispr Flow, let Claude Code update the code, then refresh and look again.

Applying the design

Once the wireframe was locked, I pointed Claude Code at the Wispr Flow website for design reference and spent about an hour iterating.

Typography: Fraunces for headings (editorial, confident) and Figtree for body (clean, readable). 2 fonts, that's it.

Color palette: 7 brand colors with names that match the vibe (lumen, vast, dawn, fathom, glow, pulse), cream backgrounds and alternating dark and light sections.

Motion: Claude Code installed Framer Motion and built 4 UI components: ScrollReveal (sections fade in on scroll), AnimatedButton (hover effects), BookingModal (a placeholder for Day 3) and Marquee (a scroller of SF neighborhoods). The before/after cards got spring physics and a tilt that follows the mouse.

One hiccup was a jitter in the SF neighborhoods marquee, tiny stutters every few seconds because the browser was recalculating layout mid-animation. Claude Code locked down the element sizing and pushed the animation onto the GPU. I came back to that thing multiple times throughout the day.

Mobile pass

The last task was testing on a simulated phone screen and fixing the major layout issues: headlines too wide, cards overlapping wrong and spacing that felt cramped. It worked on mobile, with the pixel-perfect polish still to come.

End of Day 2: From an empty project folder to a fully designed, interactive marketing site in about 6.5 hours, with 14 section components built and styled, a full design system, Framer Motion animations, interactive cards, a pill-based tabbed interface, a smooth-scrolling marquee and a first mobile pass.

Day 3: Booking & intake

Tools: Claude Code (Opus 4.6), VS Code, Playwright MCP, Notion MCP, Cal.com, Stripe, Vercel CLI, Wispr Flow, GitHub

The goal was to go from "looks like a website" to "someone can actually book a lunch and pay for it". It was also the hardest day. My daughter woke up at 1am and didn't go back to sleep until 5am, so I started the day running on fumes.

Choosing a scheduling service

Claude Code compared 8 scheduling services, and Cal.com won for one reason: native components. Most scheduling tools give you an embedded window that loads someone else's website inside yours, and it looks cheap. Cal.com renders the booking flow as a native part of your site with real animations, so the popup opens, animates and feels like it belongs.

I connected Google Calendar, configured availability and set up the $5 Stripe reservation fee. Getting the booking modal to look right took about 90 minutes of back and forth over 3 iterations: first a custom wrapper (redundant), then Cal.com's native popup (cleaner), then a final polish on colors and styling.

The intake form pivot

The original Day 1 plan was one database where each row tracks the entire customer journey, with booking info, intake answers, session notes, prep plans, billable hours and invoice details all on one line item.

I built it, looked at it and saw way too many properties. One mega-row trying to be both the customer-facing intake and the internal session tracker was unwieldy.

So I gave my concerns to Notion AI, and it suggested splitting it into 2 databases: Sessions (an internal tracker with status, prep plans, notes and billing) and Intake Responses (the customer-facing questions, with a relation back to Sessions). Notion AI built both, ran a dummy page through them and then reviewed its own work, catching stranded properties and redundant fields on its own.

Don't be afraid to blow up your own plan mid-build. The Day 1 spec said one database and Day 3 reality said 2, and I treat a spec as a starting point that the build is allowed to change.

Going live

I bought yourailunchguy.com through Vercel for $11.25, installed the Vercel CLI, linked the project, committed everything and deployed. The first production build compiled clean. Then I updated the favicon and connected the custom domain.

End of Day 3: The site is live, with the domain purchased, Stripe payments working, Cal.com booking integrated natively, the database architecture restructured, hero and testimonial images added and mobile layouts refined. I started at 8:30am and wrapped close to 2:30pm on 3 hours of sleep.

Day 4: Payments & polish

Tools: Claude Code (Opus 4.6), VS Code, Playwright MCP, Notion MCP, Cal.com, Stripe, Zapier, Vercel CLI, Google AI Studio (Nano Banana Pro), Sketch, Wispr Flow, GitHub

Today was about 2 things: making the site feel like a real product and building the automated backend.

Hero video

Earlier in the week I shot a 30-second talking head with my buddy Morgan from Outer Instinct. Claude Code compared 6 hosting options and self-hosting won easily, since adaptive streaming from YouTube or Mux is total overkill for a 30-second clip. It compressed the original 4K file (27MB) down to 1080p (7.4MB) with ffmpeg and built a "Hey, watch ๐ŸŽฌ" button that opens an animated video popup. The video only loads on click, so it doesn't slow down the page.

Automated welcome email

I needed to close the gap between "customer books" and "customer fills out the intake form". Cal.com's free tier doesn't support custom email templates, so Claude Code researched 5 options and Zapier's free tier won. It allows 5 emails a day (more than enough for 1:1 consulting), sends from my actual Gmail address and needed zero code changes.

The full pipeline: site โ†’ Cal.com booking โ†’ Zapier trigger โ†’ Gmail sends a welcome email with the Notion intake form link.

Custom assets

For images I used Google AI Studio with Nano Banana Pro. I'd feed it an inspiration image, describe what I wanted and set it to 4K. Then download, open in Sketch, magic-wand the background, delete it, export a transparent PNG and hand it to Claude Code. That gave me a confused character for the hero section and persona characters for the "Sounds Like You" section, and it replaced the remaining placeholder images.

Positioning the confused character ate 30 minutes of classic pixel-nudging with Claude Code and Playwright screenshots. At one point the image looked blurry because the code was shrinking it, and displaying it at actual size fixed it.

Notion AI agents

This was the big one. I started with one agent that was supposed to do everything, and about an hour in it was clear that was too much for one agent.

So I split it into 2:

  1. Create Session & Location agent. Triggered when a new intake form comes in. It creates a session page, links the intake response, finds the calendar event, looks at the customer's neighborhood, picks a restaurant and updates the calendar event location.
  2. Session Prep agent. Triggered when a new session is created. It reads the full intake form, interprets the customer's pain points and builds a session prep plan with tool installation checklists, MCP recommendations and workflow suggestions tailored to that person.

I ran a full test with dummy data and watched the first agent create the session, find a restaurant in Hayes Valley and update the calendar. The second agent read every intake response and built a plan that actually made sense. Pretty cool to watch it run on its own!

One issue was that the second agent didn't trigger automatically, which became a Day 5 fix.

End of Day 4: Hero video in place, welcome email automated, custom assets generated and placed, 2 Notion AI agents built and tested and real testimonial quotes added. 8:30am to 2:45pm, and I didn't even eat lunch (the irony).

Day 5: Launch & automation

Tools: Claude Code (Opus 4.6), Conductor IDE (briefly), VS Code, Playwright MCP, Notion MCP, Notion AI Agents, Vercel CLI + Analytics, Google AI Studio, Wispr Flow, GitHub

Launch day. The plan was to finish the last few site features, build out the full post-session automation and go live, so most of the day split between polishing the site and wiring up the whole customer lifecycle with Notion AI agents.

Conductor IDE detour

I switched from VS Code to Conductor to try its worktree model (isolated branches, multiple localhost instances). The workflow felt different enough that it slowed me down, and the last day of a 5-day build wasn't the time to rewire muscle memory, so I switched back to VS Code.

Build diary dropdown

Claude Code researched 6 approaches for showing Notion content on a Next.js site and built a full comparison table. After reviewing the options I made a call to just link out to the public Notion pages. The diary pages have videos, formatted blocks, tables and callouts, and rebuilding all of that natively would have been manual work every time I published a new day. Linking out means I write the page once in Notion and it's done.

The result is a nav dropdown with each day listed, purple hover states, a triangle pointer and full mobile support.

"Send This to Your Boss" section

This section lets visitors pick what's eating their time from a dropdown, get a pre-written message justifying the lunch to their boss and copy it to the clipboard in 2 clicks.

I had Claude Code build 3 completely different prototypes in isolation: a Slack chat bubble mockup with a streaming message effect, a typewriter terminal with a retro feel and a pill grid with a folding letter metaphor. All 3 ran on the first pass. I picked the Slack mockup and took it through 8 rounds of refinement.

I used the multi-prototype approach twice on Day 5 (here and on the hero video). Build 3 versions of a feature instead of one, pick the best and refine from there. It beats committing to a single direction and trying to fix it.

The final section has a dropdown with 9 options, pre-written templates under 100 words each, a ChatGPT-style typing effect (1-3 words at a time) and a copy button with a purple confirmation popup.

SEO and analytics

I did a full pass on SEO and on discoverability by AI search tools. 8 files were created or modified and the build passed first try. That covered a sitemap, a robots file, overhauled meta tags, structured data across 5 schema types and AI-specific files so LLM-powered search engines (ChatGPT, Perplexity, Claude) can read the site. It also lists all 15 SF neighborhoods so the site matches local searches.

The site explicitly allows all AI crawlers, since for a service about helping people use AI, showing up in AI search results is the whole point.

Vercel Analytics was free, took 2 lines of code and needed no separate login.

For asset cleanup I inventoried all 83 files, renamed 9 and deleted 63, leaving 18.

The 90-minute agent debugging saga

The session prep agent from Day 4 still wasn't triggering automatically. I spent about 90 minutes going back and forth with Notion AI trying to debug the trigger, checking permissions, checking the config and re-submitting test data, and nothing worked.

It turned out I was using database agents instead of custom agents. They have different capabilities, and the trigger behavior I needed only works with custom agents. The rebuild took about 15 minutes after 90 minutes of debugging.

The full customer lifecycle

With the agent architecture sorted, I built the complete post-session workflow.

Custom AI meeting notes: I wrote a custom instruction guide for session transcripts. When I record a session, the summary pulls out a session overview, tools installed, workflows built, customer Q&A, key moments and next steps.

Post-session follow-up email agent: Triggers on the "Session Complete" status. It reads the intake form, prep plan, session notes and transcript, then drafts a personalized Gmail email listing every tool installed (with URLs), a numbered list of the workflows with step-by-step instructions and 2-3 next steps. It also appends the email to the session page.

Post-session thank you agent: Fires about a week later. It drafts a reply on the original follow-up thread with a casual check-in and 2 promo codes ($20 off a next session and $20 off for a friend), and it references the specific workflow the customer built to keep it personal.

Session status flow: 7 lifecycle stages from Booked to Fully Wrapped, with invoice tracking on a separate property.

Restaurant selection: A lookup table that maps each SF neighborhood to a default restaurant where you order at the counter, seat yourself and there's plenty of room and a relaxed vibe. I hard-coded the picks instead of making them dynamic because it's simpler, more reliable and I know the spots.

End-to-end test

I ran the complete pipeline with dummy data, with Claude Code generating fake intake form submissions and dummy transcripts.

  1. Claude Code submits a dummy intake form
  2. The intake agent reads the submission, finds the calendar event, selects a restaurant, updates the calendar and creates the session page
  3. The prep agent triggers automatically and builds a tailored session prep plan
  4. I add a dummy transcript simulating a recorded session
  5. I set the status to "Session Complete"
  6. The follow-up agent drafts a personalized email with tools, workflows and next steps, then appends it to the session page
  7. The thank you agent drafts a reply with a check-in and the promo codes

All of it worked. The agents read each other's output, built on that context and produced emails that referenced the specific tools and workflows from the session.

End of Day 5: Build diary dropdown, "Send This to Your Boss" section, hero video, full SEO pass, Vercel Analytics, asset cleanup (83 files down to 18), 4 custom Notion AI agents running the full customer lifecycle, the restaurant selection table, an end-to-end test of the workflow, the agent architecture fix (database โ†’ custom agents), and the domain live with the site public.

What I built in 5 days

  • A fully designed, responsive marketing site on Next.js 16 with TypeScript and Tailwind CSS 4
  • Framer Motion animations throughout (scroll reveals, spring physics, hover effects, streaming text)
  • A Cal.com booking flow with native React components and a Stripe payment ($5 reservation)
  • A Zapier-automated welcome email with the Notion intake form link
  • 2 related Notion databases (Sessions and Intake Responses) tracking the full customer lifecycle
  • 4 custom Notion AI agents for intake processing, session prep, the follow-up email and the thank you email
  • Custom AI meeting notes instructions for session transcript summaries
  • A restaurant selection system mapped to SF neighborhoods
  • Full SEO setup (sitemap, structured data, meta tags, AI crawler discoverability)
  • Vercel Analytics and Speed Insights
  • Custom-generated assets from Google AI Studio
  • A 30-second hero video, self-hosted and compressed
  • An interactive "Send This to Your Boss" section with 9 pre-written templates
  • A build diary dropdown linking to public Notion pages
  • The domain purchased, SSL verified and multiple production deploys

How I actually used AI

The mix changed every day, but the pattern stayed the same: give AI the full context, let it do what it's good at and make the decisions myself.

Day 1 was Notion AI for content and strategy (name, voice, audience, copy and pricing), with AI as a thinking partner.

Day 2 was Claude Code for building. The Notion MCP fed it all the planning docs and the Playwright MCP let it see what it built, so the loop was read the context, write the code, look at the result and adjust.

Day 3 split between Claude Code for integrations and Notion AI for the database architecture. I ran them in parallel, working on the intake form with Notion AI while Claude Code researched scheduling services.

Day 4 added Google AI Studio for assets, which made it 3 AI tools for 3 different problems: Claude Code for the frontend, Notion AI for the agents and Nano Banana Pro for visuals.

Day 5 is where it came together. The Notion AI agents went from something I was configuring to the part of the product that runs on its own. Someone books and fills out a form, and by the time I sit down for lunch I've got a tailored prep plan I didn't write, a restaurant I didn't pick and a calendar event I didn't update. After the session the follow-up email drafts itself, and a week later the thank you goes out with the referral codes. I just review and hit send.

5 days earlier this was a blank Notion page, and now you can book a lunch.