DEV Community: Axagen

Codex vs Claude Code: Same One-Shot Prompt, Two Websites

Axagen — Fri, 16 Oct 2026 07:53:12 +0000

TL;DR: I wrote one detailed prompt, gave the exact same prompt to ChatGPT Codex and Claude Code in VS Code, and let each build a shop website in one shot. No planning step, no follow-up fixes to the design. Codex finished first but its preview didn't open at first. Claude took longer and gave me a more interactive site with more images. I liked Claude's version more, so I'm continuing with it.

Why I ran this

A website from an agency or freelancer can easily cost hundreds or thousands of euros. I wanted to see how far one prompt gets you today, with no coding and no designer. Not a finished production site, just the base: layout, design system, components and structure that I can later turn into a real shop.

The rule was one-shot prompting. One prompt in, look at what comes out. Of course a working shop needs tweaking after that, but the point was to compare the raw first output.

The setup

Gemini 3 to write the prompt
VS Code with two agents open side by side: Claude Code and ChatGPT Codex
Two separate empty project folders, one per agent, so their files don't mix
The target: a modern, clean, professional-looking shop that sells four battery-powered garden tools
Later destination: WordPress

Why WordPress? I'm in Estonia, and local shipping methods usually exist as WordPress plugins (or on platforms that have been around long enough). So whatever the AI builds has to be something I can move into WordPress later.

Step 1: Steal the prompt style, not the prompt

There are lots of website prompts shared online. One I found was made with Gemini 3.1 Pro and the site it produced looked really good. The prompt itself was the impressive part: very specific, very detailed.

I didn't want a copy of that website. So I pasted that prompt into Gemini as an example and told it to learn how the prompt is written, not to reuse it. I only have Gemini 3, not 3.1 Pro, but for writing a prompt that didn't matter.

Then I described what I actually needed: a base for a website selling four battery garden tools, modern, easy to understand, clean, but professional.

Gemini came back with a full prompt: a high-end, "creative technology" direction with core design rules, typography, components and the rest. That became the single prompt for both agents.

Step 2: Same prompt, two agents

In VS Code I created a new project folder for Claude and told it, roughly: build me a new website from the prompt I'm about to give you, it has to be usable as a WordPress site later, and I want a preview. Then I pasted the Gemini prompt.

For Codex I did exactly the same with a second folder. Only the file location changed.

One thing I normally do with Claude is hit the plan button first, so it lays out what it's going to build and I can say yes or no to parts. This time I skipped it on purpose. I already had a detailed prompt, and I wanted a true one-shot comparison.

A small UX difference while they worked: Codex streams all its text right in the panel. With Claude Code you click into the step to see what it's doing.

Step 3: What each one built

Codex finished first. Clicking the website did nothing at first: the preview didn't work. Codex then added Python, I told it to run it, and the preview came up. The hero line was "The future is silent power without fumes", with spots for products and some descriptions underneath. Clean and decent.

Claude Code was still writing a lot of files when Codex finished. When it was done it told me which file to open. At first glance it looked very similar to the Codex site, but the buttons had a nice movement on them, there were more images and more sections overall.

Then I put them side by side, Codex on the left, Claude on the right.

Results

	ChatGPT Codex	Claude Code
Prompt	Same one-shot prompt	Same one-shot prompt
Finished	First	Later, still writing when Codex was done
Preview	Didn't open at first; worked after it added Python and I told it to run	Told me which file to open, opened fine
Button movement	Different style	More interactive
Images	Fewer	More
Product view	Placeholder spots plus descriptions	Different layout and style
Bottom of the page	Different content	Different content, more stuff overall
My pick		âœ…

The honest summary: both are a usable base from one prompt. The texture, style and product view are different between the two. Claude's felt more "wow", more interactive and nicer overall. That's my taste, though; you might prefer the Codex one when you see them.

What failed (or didn't happen)

Codex's first preview failed. It needed a follow-up from me to run it. Strictly speaking, that's the one place the "one shot" wasn't fully one shot.
Neither is a real shop yet. These are design bases with spots for products, not finished working stores. A full working site still needs tweaking.
No WordPress yet. I asked for something that can become a WordPress site, but I didn't convert or test anything in WordPress in this run.
I skipped Claude's planning step. That kept the test fair, but it's not how I'd normally work. With planning, both might have done better.
One run, one prompt, one niche. This is a single comparison, not a benchmark. A different prompt or a second run could flip the result.

What I'd tell another builder

Spend your effort on the prompt. Using a strong example prompt as a teacher, and having another model write the real prompt for your niche, did most of the heavy lifting here.
Give each agent its own folder when you compare them, or they'll trip over each other's files.
Don't judge on "finished first". Codex was faster, but the preview didn't work until it got another nudge.
Decide early where the site has to end up (WordPress for me, because of local shipping plugins) and say it in the prompt.

I'm continuing with Claude. The next step is building an actual website with an actual product that can really run on WordPress.

Timestamps

Written up from my video with AI help; the test and the opinions are mine.

Full run on video (my channel): https://www.youtube.com/watch?v=AR-A7Ro1te4

I Let an AI Bot Trade Real Money for a Week – Results ($50 Live + $20k Paper)

Axagen — Tue, 13 Oct 2026 21:03:18 +0000

Not financial advice. Small money, short window. This is an experiment log, not a strategy recommendation.

TL;DR: I gave Grok Bot $50 of real money on Alpaca for about a week, plus $10,000 in paper stocks and $10,000 in paper crypto. Real money ended down 38 cents after fees. Paper stocks lost $88. Paper crypto was down about 2–2.7%. It's not a get-rich story, and I'm keeping the experiment running.

The setup

I started with a simple question: can Grok Bot make me money? Trading was the obvious first test.

Attempt 1: eToro (blocked). I already use eToro, and I had some leftover money there that I wasn't scared to lose. The plan was for Grok Bot to open Chrome on its own computer screen, log in and trade by hand. eToro kept logging out, and then kept blocking the session as suspicious activity, so I'm guessing it doesn't accept bots. After two or three days of errors I dropped the idea.

Attempt 2: Alpaca (API). I looked for trading platforms that allow trading through an API and found Alpaca. Generating an API key is easy, and the setup was simple, especially when you hand it to the AI.

I ran two tracks:

Live: $50 of real money, fully handed to the AI. It ran from 24 September, exactly one week before I recorded on 1 October.
Paper: $10,000 in stocks and $10,000 in crypto (fake money), running since 11 September.

The rules I gave it. The easy move would be to buy an S&P 500 fund and hold it. I wanted something more active, to see what it can actually do:

Sell daily if possible.
Hold overnight only if there's a very good reason.
Sell everything by Friday.
Research the market before deciding what to buy.

Early on it also had a rule not to trade when it couldn't find anything good. It didn't trade at all, so I told it to actually trade. Maybe those conditions were too extreme, and that's on me. With $50, a good day might make nine cents, so this was never about the money.

Tools and cost

Item	What I saw
Agent	Grok Bot (prompts via chat; there's also a live voice-talk mode on the phone that I skipped because it uses up credits)
Broker	Alpaca, trading through its API
Paper stock commission	0%
Paper crypto fee	0.25% per trade
Live trading fees	Included in the −$0.38 result
Funding (EU)	Alpaca isn't in the EU, so I sent the money through a UK bank account. It took 2–3 days and was a bit of a hassle, but not that bad

Results

Account	Start	Period	Result	What it traded
Live (Alpaca)	$50	24 Sep – 1 Oct	−$0.38 net after fees (dividends +$0.11)	SPY (S&P 500), stop-losses ~1% below buy price
Paper stocks	$10,000	since 11 Sep	−$88 (the bot reported roughly −1%)	S&P 500 index funds
Paper crypto	$10,000	since 11 Sep	−2.7% at recording	Bitcoin and Ethereum only

Live, day by day: It didn't trade at first. Friday: bought SPY (fee included). Monday: bought more SPY. Tuesday and Wednesday: losses. At the end it was still holding SPY shares. Losses came from the stock moves and fees, and dividends brought in 11 cents.

Paper stocks: Friday was up $34, and every other day was a loss. Same funds and same period as the live account, so similar results.

Paper crypto: It had bigger daily losses than stocks.

What failed

Screen-login trading on eToro. It got blocked as suspicious activity, and I lost 2–3 days to errors.
The "don't trade unless it's good" rule. The bot sat idle, so I forced it to trade, and some of the loss probably comes from that pressure.
Timing. The bot's main lesson was: "deploy gradually, not all at once near a high."
The bot's own explanation. When I asked where the losses came from, it said mainly its own decisions plus the thin account, with bad luck as the smaller part. I can't blame the AI completely, though. If it's a bad week for the market, it's a bad week for the bot too.
Stop-losses. Every stop sat about 1% below the buy price, and the bot said the Monday sale was the stop working as intended.

What I'd change

Run it longer. One week of real money and a few weeks of paper isn't enough to decide anything. I plan to keep it going for another month or two.
Add a weekly learning session. Every Friday after the market closes, the bot reviews everything it did, works out its mistakes and what went right, writes those lessons down (in its notes or a library) and gives me a weekly recap of what it learned and what to do better. By the end of the recording it had already added this session to itself.
I'm still deciding whether it's worth keeping long term.

Timestamps

Not financial advice. This is $50 of real money over one week. Don't read any of it as a reason to trade.

Full run on video (my channel): https://www.youtube.com/watch?v=sprcALCNoW0

DEV Community: Axagen

Codex vs Claude Code: Same One-Shot Prompt, Two Websites

Why I ran this

The setup

Step 1: Steal the prompt style, not the prompt

Step 2: Same prompt, two agents

Step 3: What each one built

Results

What failed (or didn't happen)

What I'd tell another builder

Next

Timestamps

I Let an AI Bot Trade Real Money for a Week – Results ($50 Live + $20k Paper)

The setup

Tools and cost

Results

What failed

What I'd change

Timestamps