Back

Your First Side Project With AI, for People Who Don't Code

Aug 23

Your First Side Project With AI, for People Who Don't Code

Open any tech forum this week and it's the same conversation. V2EX, the biggest Chinese-language dev forum, had three vibe-coding threads trending in a single day, one of them running to 65 replies. The Hacker News front page had a project about a new way of coding with AI at 378 points. Japan's Zenn is deep into advanced Claude techniques. The Chinese-speaking crowd is arguing about the thing itself, the English-speaking crowd about workflows, the Japanese crowd about power-user tricks.

All of these conversations share one assumption: the reader writes code.

This post is for everyone else. Designers, marketers, teachers, shop owners. You have an idea and you're willing to put in an evening, but a black terminal window full of white text makes you close the tab. I want to translate this week's noise into something you can actually use.

What vibe coding actually is

One sentence: you describe what you want in plain language, the AI writes the code, you look at the result, and you tell it what to change — also in plain language.

"The button is too small." "Make it green." "I want to export this to Excel." The AI understands all of these and just does them. At no point do you need to read a line of code.

In 2026 this genuinely works. But I want to draw the boundary of "works" precisely, because a badly drawn boundary is where most of the disappointment comes from:

Works: a single-page tool (a calculator, a random picker, a word counter), a form that collects real responses, a mockup of an interface you can show people. Build it, use it yourself, hand it to ten friends — that tier is a weekend project now.

Don't start here: user accounts, credit card payments, anything meant to survive thousands of simultaneous users. AI will produce something that looks like those things, but it won't hold up, and when it breaks you won't know where to start. Pick a different first project.

Below are three routes you can start for free, each with a concrete first 30 minutes. Every price and plan detail here is verified (checked 2026-08-23, some figures carried over from our early-August audits); where I couldn't confirm something, I say so instead of guessing.

Route one: talk to Claude until a web page falls out

This is vibe coding proper.

Your first 30 minutes:

  1. Go to claude.ai and sign up with an email or Google account. Free, no credit card.
  2. You may have heard of "Claude Code." That's a terminal tool for engineers, it is not included in the free plan, and you don't need it. The chat window on the website is enough.
  3. Type a description of what you want. Here's a template you can copy and adapt:

Build me a web page: I paste in a block of text, and it counts the words and estimates reading time. Make it a single HTML file with no external services — I want to save it on my own computer and use it there.

  1. Here's what actually appears, because it surprises people: Claude does not dump code into the chat. A panel opens beside the conversation showing your page already working — paste some text in and watch it count, right there. Claude calls this panel an Artifact. The code exists, but it's tucked behind a "Code" view at the top of that panel, and you never need to open it.
  2. Get the file onto your own computer. In that panel, look for the download button (hover over the icons in the panel's corner until one says something like "Download"), click it, and save the file — that's it, skip to step 6. Only if your reply came as a plain code block inside the chat instead of a panel: press the copy button on that block, then on Windows open Notepad, paste, Save As, name it tool.html, set the file type to "All files"; on Mac open TextEdit, first go to the Format menu and click "Make Plain Text" (skip this and the file comes out broken), paste, save as tool.html. Whatever you do, don't try to select the code with your mouse — you'll drag in surrounding sentences and save a broken file.
  3. Find the file and double-click it. Your browser opens, and what's on screen is your tool. A working one.

How to iterate: go back to the chat and say what you don't like in plain words. The page in the panel updates itself; download it again and replace your saved copy. If you're on the copy-paste route instead, ask for "the complete updated file" every time — always the full file, never a snippet, because you have no way of knowing where a snippet is supposed to go.

The honest limits: whether the free plan has a usage cap, and what it is, I could not confirm — the official pricing page told me "free, $0" and nothing about specific limits, so I won't guess. If mid-conversation it asks you to wait or upgrade, that's what you've hit. If you get hooked: Pro is US$17/month billed yearly or US$20/month billed monthly, and Claude Code starts at Pro too. For a first weekend, start free.

Route two: Tally, something people can use today

If your idea boils down to "collect responses from people," you don't need to generate any code at all. Pre-orders for a shop, sign-ups for a class, intake forms for freelance work — all this shape.

Your first 30 minutes:

  1. Go to tally.so and create a free account.
  2. Make a new form and type out your fields: name, contact, the questions you actually want answered.
  3. Hit publish and you get a link.
  4. Paste that link into your Instagram bio, your group chat, your page.

Thirty minutes in, you have a thing that is genuinely collecting data. I mean it when I say this counts as a side project: a pre-order form with real responses beats any half-finished prototype sitting on a hard drive.

What the free plan actually includes (checked against the official pricing page on 2026-08-23): unlimited forms and unlimited responses within fair use, plus conditional logic, file uploads, e-signatures, payment collection, and Google Sheets/Notion integrations — all free. The one cost: your forms carry Tally branding. Removing the branding is what you pay for, on Pro at US$24/month (annual billing gets you two months off). Note the direction: free keeps the branding, paid removes it. Some articles out there state this backwards.

Route three: Google Stitch, draw the screen first

Some people get stuck one step earlier: the idea hasn't taken shape yet, so there's nothing concrete to describe to an AI. Then draw the screen first.

Your first 30 minutes:

  1. Go to stitch.withgoogle.com and sign in with a Google account. Free.
  2. Describe the screen in your head in one sentence — "a phone app home screen for logging daily shop revenue, big number up top."
  3. It generates an interface design. Don't like it? Rephrase and generate again.
  4. Take the result to a friend, or hand it to Claude from route one: "build this screen as an HTML file."

The honest limits: Stitch is a Google Labs experiment. It's currently free and the official page shows no pricing at all, but there is a usage cap, Google hasn't published the numbers, and the figures floating around secondhand contradict each other — so I'm citing none of them. "Experiment" also means it could change or disappear without much notice; don't build a long-term dependency on it. We wrote up a hands-on review if you want to go deeper.

Where you will get stuck: three honest warnings

All three routes share a property: your thing lives on your computer or on someone else's platform. Take one more step outward and you hit these three walls. Knowing where they are makes the collision less scary.

Wall one: deployment, i.e. putting it on the internet

Your HTML file runs fine on your machine, but you want to give people a URL. That step is called deployment, it's where the terminal shows up, and it's where most non-engineers quit.

The honest state of things (from our provider-by-provider audit in early August): if you want "free, always awake, with a real database," the only remaining combination is Cloudflare's — and its setup is not friendly to non-engineers. Render still has a genuinely free plan, at the cost of falling asleep after 15 minutes of no traffic, so your next visitor waits while it wakes up. Railway's trial credit is a one-time US$5 for 30 days, and the honest number after that is US$10–20/month. Fly.io hasn't had a free tier since October 2024 — don't trust older articles. The honest number overall: an always-on service with its own database runs US$5–10/month.

My advice for a first project: don't deploy. Send the single HTML file to friends over cloud storage or chat, with one caveat that will save you a confused evening: this works on computers, not phones. On a computer, your friend double-clicks the file and it opens in their browser. On a phone, tapping an .html file in a chat app or a cloud-drive app opens a document preview where the tool doesn't run — they'll see raw code or a dead page, tell you "it won't open," and you'll both think the thing is broken when it isn't. So say "open this on a computer" when you send it, and for the phone-only people, demo it in person or send a screen recording instead. Ugly, but zero cost and zero risk. When you're sure the thing deserves a real URL anyone can tap from a phone, read our free hosting audit and free tools roundup — both are checked price by price.

Wall two: the domain, the one unavoidable cost

If you want your own address (yourname.com), that costs money — roughly US$12/year, and it's the only truly unavoidable cost in the whole side-project stack. A verified example: Porkbun sells .com registration and renewal at US$11.08/year, including WHOIS privacy (so your personal details aren't publicly searchable) and an auto-renewing SSL certificate (the padlock next to the address).

Skipping this is completely fine too: use the address the platform gives you. Tally's publish link, a hosting platform's subdomain — for a first project these are all you need.

Wall three: your data will disappear

Three common ways. One: the free tier itself dies — Fly.io is the precedent; your project lives on someone's free plan, and landlords change the rules. Two: the service sleeps or deletes data on its end. Three, the most common: you overwrite your only copy yourself.

The countermeasures are unglamorous and effective. Save a copy of every complete file the AI gives you, so a bad change can be rolled back. Export form responses regularly — Tally's free plan connects to Google Sheets, which is automatic backup. Anything you care about should exist in more than one place.

How to ask for help when you're stuck

Move one, always: copy the entire error message, paste it back to the AI, and lead with "I'm not an engineer, walk me through this step by step." The error can look like gibberish to you; the AI reads it fine. For visual problems, screenshot and describe. When asking anyone for help, cover three things: what I did, what I expected, what I actually saw. "It's broken" alone gets you nothing.

The last thing: building it is not the same as keeping it alive

This is the paragraph I most wanted to write. The AI helped you build the thing. It will not maintain the thing.

Three months from now: a browser update changes some behavior and your page looks off; a service you connected changes its rules; you want to add one small feature, but you never understood the original code and you won't start now. This isn't your failure — an engineer's project rots the same way after three months of neglect; they just repair faster.

I have two countermeasures. First, build small: one file, few external services. Fewer moving parts, fewer things that can break. Second, write your requirements out in plain language in a document and keep it. When the thing breaks or you want a new version, hand the same document back to an AI and have it rebuild from scratch. For a non-engineer, rebuilding is almost always faster than repairing.

Find 30 minutes this weekend and pick one of the three routes. What you make will probably be small, a bit ugly, and used by exactly one person. That's fine. First ones always look like that.

This post comes from SideShip, a place for sharing side projects. If you have one half-built, or finished and unseen, post it here — three required fields, no email verification, two minutes. When you publish, I'll personally leave the first piece of feedback.