"hi, make me a business"

Alright,

nobody hires someone on Monday and says “make me a business.”

But you’re basically saying that when you open up an AI tool and scramble together “build me a booking app for dog groomers.” You’ve just taken on works around the clock on your objective, never complains, but knows nothing about your business or why it exists. Then you sat it down at an empty desk with no brief, no accounts, no brand guide and no process, and asked it to get on with it.

So here’s the onboarding package for your agents, and every kit in it is free!

In today’s email: why apps built prompt-by-prompt collapse around feature five, the four things to hand your agent before your first prompt, what day one actually costs in dollars, and the free-tier rule that kills apps the week they start making money.

In my experience, a successful end to end buildout of a vibe coded app is massively dependent on the setup and ‘onboarding’ you give it.

A new hire gets some paperwork, or more like a notion file these days. Yours is a plain file called architecture.md, and it says what the app is, what information it stores, who’s allowed to see what, and what order to build things in.

Writing one of these normally requires precisely the technical vocabulary you don’t have yet, which is a bit like being told you can have the map as soon as you know the way.

So don’t write it. Get interviewed for it. Paste our interview prompt into Claude or ChatGPT and let it ask you about twelve plain-English questions, one at a time. It explains each concept in a sentence as it goes, makes the technical calls itself, and hands you the finished document. Then that document becomes the first thing your builder ever reads.

It’s free, and the page has a finished worked example for a dog groomer’s booking page so you can see what good looks like before you start. If you think better in pictures, sketch the screens first in Excalidraw (free, open source, no account needed) and paste the sketch into the interview. The AI reads it fine.

One rule matters more than what’s in the file: the brief is what you argue with, not the agent. When feature five conflicts with feature two, you don’t debate it in chat. You open the brief, decide, and update it.

You’re going to want to make sure the agents have all the tools they need.

• The builder: pick one, then stop shopping. All three land in the same place at $25/month, so the free tiers are what actually separate them. Lovable gives you 5 credits a day, roughly 2 to 4 real prompts. Bolt has the most generous free tier of the three, 1M tokens a month with a 300K daily cap. Replit sits closest to a real development setup, which makes it the one to pick if you suspect you’ll outgrow the browser. Rule: the builder matters far less than the brief you hand it. Switching in month two is cheap, a missing architecture.md is not.

• The database and logins: Supabase. Open source, and the free tier is genuinely usable: 500MB of data, 50,000 monthly users. One trap nobody mentions: the free database pauses itself after a week of no activity, which looks exactly like your app breaking. Rule: one system holding both your data and your logins beats two you have to duct-tape together.

• The money: Stripe. Don’t worry about the specifics, everyone uses Stripe and it’ll be the best bet for you.

• The safety net: GitHub. Free, and it is the undo button for your entire app. Rule: non-negotiable, especially if you can’t code. Steps below.

The math on day one: $0. Every tool above has a free tier that covers a real first version. You’ll be at $25/month inside a fortnight, and that’s the whole bill until people are actually using the thing. The full annotated version of this stack, including which free tiers pause on you, is in the Vault.

GitHub can only save versions from the moment you connect it. Wire it up now and “go back to Tuesday” becomes a real sentence. Leave it until something breaks and it isn’t.

1. Make a free account at github.com. Nothing else to configure.

2. Connect it inside your builder.

   • Lovable: Project settings → Git → GitHub, find your account under Accounts, hit Connect. It creates a private repo and syncs both ways, so every change the agent makes becomes a saved version (docs).

   • Replit: the Version Control tab in the left sidebar, then Connect to GitHub. Private repos need the $25/month Core plan (docs).

   • Bolt: connect your GitHub account in Settings, then push the project to a new or existing repo (docs).

3. Build one small thing and check a commit appears in the repo. If it doesn’t, fix that today, while you still have nothing to lose.

The version of this with every screenshot and no jargon is in the Vault.

Now the expensive part, and this is the one I would tattoo on beginners.

Every guide tells you to deploy free on Vercel. Almost none mention that Vercel’s Hobby plan is non-commercial only, on a definition broad enough to cover a freelancer you paid to build it. They can disable the deployment without notice.

Translation: the free tier ends the day you succeed. The moment your app takes a dollar you’re on the paid plan, or one enforcement email from being offline. The clean way to get an app onto a real domain is in the Vault.

You wouldn’t let a new hire redesign the logo on day one. We covered this properly last week: left unconstrained, your agent picks the most statistically probable design every time, which is why every AI-built app arrives with the same purple gradient and floating rounded cards.

The fix is the same shape as the brief. Screenshot a site whose look you’d be proud of, run it through a prompt that reverse-engineers its colours, fonts and spacing into a design.md, then drop that file into your builder’s standing-instructions slot so it applies to every screen you ever build. Fifteen minutes, free in the Vault.

Hunting grounds: Mobbin for product screens, SaaS Landing Page for landing pages. To nudge the stolen palette by eye before locking it in, tweakcn is free and built for exactly the themes these builders use.

The files are the induction. What follows is the actual job.

1. One task per session. Take the next item off your brief’s build order. Not three, however tempting. Sessions that sprawl are sessions where the early instructions quietly stop mattering to you.

2. Check the work yourself. Not “does the chat say it’s done,” because we already know what that’s worth. Click the actual button in the actual app. That 91% figure is the entire reason this step exists.

3. Two strikes, then revert. If the agent has failed to fix the same bug twice, stop prompting. Roll back to the last version that worked and describe the feature differently. A third attempt rarely works, and it usually takes something else down with it. When you can’t tell what the error even says, run it through the error message decoder before you paste it back into chat.

4. Say it twice, write it down. The second time you correct the same thing in chat, it doesn’t belong in chat. It belongs in architecture.md or design.md.

That’s the whole operating system. Everything after this is just running it again next week.

• Claude Code made Opus 5 its default on Friday, with a 1M-token context window and fast mode at $10/$50 per million tokens. More context means longer sessions before things drift, but rule 1 still stands. Longer rope is not the same as no rope.

• The open-model fight got loud. Moonshot released Kimi K3’s weights on Sunday, the first openly released 3-trillion-parameter-class model, and Anthropic broke its silence to say it has never advocated banning open-weight models. Top two stories on Hacker News. You’ll never run this on your laptop, but you benefit anyway: this is the pressure that keeps dropping the price of the tools you do use.

• Cursor doubled included usage on every plan last Tuesday. If you bounced off it on limits, the maths just changed.

• Yap: press a shortcut, talk, and your words paste wherever you were typing, including your builder’s chat box. It ships no model and makes no network calls, borrowing the speech engine already in macOS. 4MB. Needs macOS 26 or later, which is the catch.

• i-have-adhd: a skill that stops your coding agent burying the answer. Leads with the command or file path, numbers multi-step work, caps lists at five, bans “Great question!” Nearly 7,000 stars in a week, which tells you how many people were quietly annoyed.

• OmniRoute: one endpoint routing to 500+ models across 290+ providers, 90+ of them free, with automatic fallback when you hit a quota wall mid-build. For when you’ve outgrown the browser and you’re tired of paying retail per token.

Keep your instructions file short. Every builder has a slot for standing instructions, and the temptation is to stuff it with a 500-line rule pack you found on GitHub. Don’t. A study of 138 real projects found AI-written instruction files actually reduced success rates and pushed costs up over 20%, while human-written ones helped only when they were short and specific. Rules the model can hold beat rules that impress you. Ours is under 80 lines, with the reasoning for every one of them.

• How Coding Agents Fail Their Users: the 20,574-session study behind today’s issue. Skip to the failure taxonomy.

• PGSimCity: an explorable 3D city showing how Postgres actually works, on the Hacker News front page today. If you just signed up for Supabase, this is the beautiful version of what you bought.

• Create an Onboarding Plan for AI Agents: Harvard Business Review got to today’s metaphor before I did, aimed at big companies. It scales down to one person better than it has any right to.

• Ask HN: What Are You Working On?: this month’s thread, 1,152 comments of people describing what they’re building. A better idea generator than any “100 app ideas” listicle you’ll ever open.

• Decker: free, open source, runs in your browser, and makes small interactive things the way HyperCard did in 1987. On the Hacker News front page this week. There is no AI in it anywhere, which is exactly the point.

Cheers folks!

Jonathan