🧭 THIS WEEK AT BuildProven
Howdy, back from vacation and so slowly getting back into routines. still hard with summer. Few things -
Kim 3 released and looks pretty good - i use openrouter and I don’t think it’s so available yet - will look at using it in coming weeks.
ChatGPT 5.6 seems pretty good - so much so, for now I am trying to switch to codex as primary with claude as backup only. It’s taking a bit of getting used to but chatGPT token price definitely better.
the ChatGPT app I like a bit better - it separates remote connection [using your computer] into separate threads/projects better than Claude ‘dispatch’ equivalent does.
GPT is faster than Claude i think.
it’s taking a bit to get my whole setup working seamlessly either for claude or gpt.
meantime, i’ve been shipping a bunch of stuff -
starknet/my openclaw - fixed a LOT and cleaned a LOT of stuff/technical debt etc.
Claude Kit released FREE as open source - works for both claude and GPT
QA Architect and SAAS Starter Kit both updated/refreshed (security, bugs, dependencies etc)
Two new quick little projects - very much alpha stage -
wanderpage - automatic travel site creation from your travel photos
lifescape - where do you want to live after the grind?
as usual, would love it if you could share this newsletter, give feedback via the poll at the end, or best, just hit 'reply' and tell me what you need/want with AI. I will definitely reply and try to help best i can.🧰 Worth Your Click
Deploying from the Vercel CLI — the canonical page for the one-command deploy above. If the command name or sign-in flow ever changes, this is the source of truth to check instead of guessing.
Claude Code troubleshooting — the official page for when a prompt does something you didn't expect or a step won't run. Worth bookmarking before Saturday so you're not searching mid-build.
🗺️ FEATURED INSIGHT
Three weeks ago you installed Claude Code and made it change one file. Then a folder of 200 notes. Then you learned Codex does roughly the same job. This is where it was heading: this Saturday you build one small thing that's actually yours, and by 4pm it's live at a URL you can open.
Not an app for work. Not a demo. The kind of tool you've meant to find for years and never quite trusted an app store one with. You're going to build it with five prompts. The code stays Claude's job. Yours is the design: what it holds, what it does, what "done" looks like.
Here's the whole plan, start to live URL.
Pick one (three specs — not toy examples)
Don't build a TODO app. Everyone builds a TODO app, uses it twice, and forgets it. Pick a project that answers a question you actually ask yourself. Here are three that do, each one small enough to finish in a day.
Spec A — Weekly habit tracker.
What it does: shows the current week (Mon-Sun) with your handful of habits down the side. You tap a box to mark a habit done for that day. It shows your current streak per habit.
What it holds: a short list of habits (you set them) and, for each one, which days you ticked. Stored in the browser, so it's private to your device — no login, no account.
Done looks like: you open the URL Monday morning, see this week's grid empty, tick "walked 30 min" on the days you did it, and the streak counter goes up. Last week's ticks are still there when you scroll back.
Spec B — Recipe inbox.
What it does: one box where you paste a recipe (or a link and a few notes), hit save, and it lands in a searchable list. Click any entry to read the full thing. Search by name or ingredient.
What it holds: each recipe's title, the full text you pasted, and a couple of tags you add ("quick", "vegetarian"). Stored in the browser.
Done looks like: you paste the curry recipe your sister texted you, tag it "quick", and next Tuesday you search "curry" and it's there in one click — instead of scrolling three months of messages.
Spec C — Reading log.
What it does: log a book or article with a one-line "why I'm reading it" and a status (want-to-read / reading / done). When you finish one, add a two-sentence takeaway. Shows a running list of what you've finished this year.
What it holds: title, author, status, your note, your takeaway. Stored in the browser.
Done looks like: you mark "finished" on the book you just read, type the one thing you want to remember from it, and your "done in 2026" count ticks to 12. In December you have a real list instead of a vague sense you read some things.
All three store their data in your browser, on purpose: no login to build, no database to set up, private to you from the first minute. The trade-off you're accepting: the data lives on the one device and browser you use it in. Right call for a private project you build today; when you outgrow it, that's a different Saturday.
Pick the one you'd genuinely open next week. The build is identical for all three, only the words in Prompt 2 change.
The five prompts, in order
Open Terminal. Make a fresh folder for the project and move into it, then start Claude Code — the same two-line move you've done for three weeks:
mkdir ~/habit-tracker && cd ~/habit-tracker
claude(Name the folder for the project you picked — ~/recipe-inbox, ~/reading-log.) Now type these five prompts one at a time. Read each answer before you send the next.
Prompt 1 — set the ground rules before it builds anything.
I'm not a developer. I want to build a small single-page web tool that runs in a browser with no login and no server — it should store its data in the browser itself (localStorage). Keep it to plain HTML, CSS and JavaScript in as few files as possible. Before you write anything, tell me in plain English how this will work and what files you'll create. Don't write code yet.This is the seatbelt, same as week one. You're pinning the shape — one page, no server, no login — before it starts, so it can't wander off and build something that needs a database you'd have to run.
Prompt 2 — hand it the spec you picked.
Good. Build this: a weekly habit tracker. It shows the current week Monday to Sunday with my habits down the left side. I tap a box to mark a habit done for a day. It shows my current streak for each habit. I can add and remove habits. Everything saves in the browser so it's there when I come back. Make it clean and easy to read on a phone. Build it now.Describe instead the recipe or reading log if you chose those from the above specs. The more specific you are here, the less you steer later. Vague in, vague out.
Prompt 3 — see it before you trust it.
How do I open this and try it right now on my own machine? Give me the exact command or the exact steps, in plain English, assuming I've never done this.It'll tell you how to open the file in your browser (often as simple as open index.html). Open it. Click around. This is the moment the thing becomes real — a tool running on your screen that didn't exist this morning.
Prompt 4 — fix what's wrong by describing it, not coding it.
I tried it. Here's what's off: [describe exactly what you saw — "the streak counter didn't go up when I ticked two days", "it looks cramped on my phone", "the habits disappeared when I refreshed"]. Fix those, then tell me what you changed and how to test that it's fixed.You are not debugging code. You're describing what you saw, the way you'd tell a colleague "the total's wrong on the second tab." That's the whole skill of this issue: you judge the result, Claude fixes the cause.
Prompt 5 — lock it in.
This is right. Now write me a short plain-English note in a file called README.md explaining what this tool does and how to run it, so I understand
it in six months. Then tell me the single simplest way to put this on the internet as a private URL only I have.The README is you keeping the receipts on your own project. The second half sets up the deploy — which is the last step below.
The two places it goes wrong (and the exact fix)
Every first build hits at least one of these. My first one broke in exactly the second way below — the data vanished on every refresh and I assumed I'd done something wrong, when it was just a missing instruction. None of these needs you to open the code. Each one is a prompt you paste.
1. It builds something that needs a server. You asked for "no login, no server" in Prompt 1, but on a bigger project the model sometimes reaches for a database anyway, and then "how do I run it" turns into a rabbit hole of installs. The tell: the run instructions in Prompt 3 involve installing something or starting a "server". Fix it by pulling it back:
Stop — this is more complicated than I need. Rebuild it as a single HTML file that I can open by double-clicking it, with all the data saved in the browser
(localStorage). No server, no database, no install. Same features, simpler.2. The data vanishes on refresh. You tick your habits, refresh the page, and it's blank. This means it never actually saved — it held the data in memory only. Very common on a first pass. Fix:
The data disappears when I refresh the page. It should persist. Save every change to the browser's localStorage and load it back when the page opens. Then tell me how to test that a refresh keeps my data.Notice the shape of all fixes: you describe the symptom in plain words and hand the cause back to Claude. Same move as the wrong-vs-right from week one — "clean up my notes" fails, "add a one-line summary to each file, change nothing else" works. Here it's "fix it" versus "the data disappears when I refresh; make it persist and tell me how to test it." Specific symptom, checkable fix.
Deploy it — one command to a live URL
You've got a working file. Last step: put it on the internet so it's a real URL, not a file on your laptop. The dimension you want is simple — one command turns the folder into a public web address. Two paths:
Vercel is the common one. You install its command-line tool once, then from your project folder you run one deploy command and it hands you back a live
https://URL in under a minute. Ask Claude Code to walk you through it:"Walk me through deploying this folder to Vercel from the command line, step by step, assuming I've never used it. Stop after each step and wait for me."It'll install the tool, sign you in through the browser, and run the deploy. The exact command name and flags shift over time, so let it read the current steps rather than trusting a flag I write today.If you'd rather not create an account anywhere, any static-file host does the job — the shape is always the same: point it at your folder, receive a URL. The one-command-to-a-public-URL pattern is what you're after, whichever host you use.
Either way, by early afternoon you paste the URL into your phone and there's your habit tracker, live. That's the finish line: a real tool, built from a spec you wrote, running at an address you own.