💻 Programming

[late sphinx]

thanks! I've been trying to get as much as I can out of the cursor free token period. I have a new version with more fireworks shapes, a new tank game... it's been great getting to play with frameworks like phaser without needing to learn them first for prototyping. I would invest the time to learn the tools if this was a production app but sometimes just seeing if they are the right tools takes a long time and this totally shortcuts it.

[median wasp]

I built (and shipped!) a production eBay automation in two weeks—with GPT-5 as my senior engineer

TL;DR: I’m not a software developer. In two weeks, working evenings with GPT-5 as a kind of senior engineer/pair-programmer, I built a real FastAPI service that scans eBay for arbitrage opportunities, stores “opportunities,” lets me save clean listing drafts, and—after I physically receive an item—creates & publishes the listing via the eBay Sell APIs. It runs on Render, code lives on GitHub, and all data is simple JSON. Along the way we solved OAuth scope tangles, policy/marketplace mismatches, the infamous “25007 shipping option” error, and a tricky condition-mapping issue, plus added admin tools and a local “draft store” for long lead times.

---

What GPT-5 built for me:

A small FastAPI web service that:

POST /scan: queries the eBay Browse API with an application token, scores results with a profit engine, and persists only “opportunities” (title, snapshot price, link).

GET /: a simple HTML page showing the opportunity feed (images, snapshot price, margin, actions).

“Save Draft” button per opportunity that writes a clean, minimal listing draft (JSON) to my local “Draft Store” on Windows, so I can come back days later after the parcel arrives.

POST /confirm-receipt: turns a saved draft into a published eBay listing using Inventory + Offer APIs (and Fulfillment for shipments later). It uses a user access token minted from a stored refresh token.

A tiny admin menu: list/inspect/delete drafts and opportunities, plus optional upload/restore of a draft.

A conservative design:

No auto-buy. The service surfaces links + a price snapshot; I buy manually on eBay.

Simple storage: flat JSON files (opportunities/…, drafts/…) and a lightweight template.

A profit engine with environment-driven costs (fees, shipping in/out, packaging, return buffer, VAT), computing net profit and margin with clear “reason” strings for debugging.

---

What surprised me most as a non-developer:

The “coding” part was only 50% of the work. The other 50% was API product knowledge: scopes, marketplaces, Business Policies, condition IDs, sell privileges, tax compliance, and what belongs in an Offer body vs. a Policy. GPT-5 was priceless here—it kept nudging me to the correct docs model (e.g., “shipping lives in the policy, don’t push shippingOptions into Offer”).

The feedback loop was fast: copy a curl command, look at the JSON error, paste the snippet to GPT-5, ship a tiny patch, repeat.

Plain JSON + files was more than enough. I didn’t need a database for v1.

---

Why this matters (my take on the societal angle):

I’m a non-programmer who built and shipped a working, policy-aware automation in two weeks. If this scales, we’ll see:

Entrepreneurial acceleration: one founder + one AI can validate ideas at the speed that once required a small engineering team.

Shift in scarce skills: Knowing the domain (e.g., e-commerce compliance) becomes as valuable as typing code. AI narrows the gap between “I have a process in my head” and “it runs in production.”

Compliance by construction: Paradoxically, more automation may produce safer systems, because a good assistant keeps you inside platform rules (scopes, policies, rate limits) and documents why.

Open tooling ecosystems: With Render/GitHub and a handful of libraries, deployment isn’t a cliff anymore. The barrier is mostly clarity and patience—not money.

I don’t think this eliminates professional developers; it reassigns them to deeper problems. But it does mean thousands of “small automations” can now be built by the people who understand the workflows best—with AI as the senior engineer who never sleeps.

If you want to reproduce this, start simple: HTML list + “Save Draft,” flat files, one marketplace, one account. Keep a “who am I” endpoint to verify tokens. Add features only when a real error forces you to learn the next layer (condition policies, privileges, etc.). That’s how I got to my first fully automated listing.

If this helped, feel free to reuse the structure and adapt it to your market or platform. And if you’re stuck on a 25007 or 25019 at 2 a.m.—I’ve been there. GPT-5 has, too.

[finite sparrow]

Has anyone faced an irritative bug in canvas when programing in c/c++ or similar and using c-style string like "some string\n". When AI updates canvas it always converts \n to a real line break. When asking them to use escaped \ \n it put escape in a final code as well. Any possible solution?

[obtuse steeple]

Here's the continuation of my experiments with ChatGPT! This time, the video presents an experiment integrating ChatGPT-5 into a 3D environment to test the interaction capabilities of this type of AI by playing chess!
Although the goal isn't just to demonstrate the AI's mastery of the game: chess serves as a visual aid to illustrate the new perspectives offered by this type of interface.

I remind you that no API is used: the connection is made via the same channel as a traditional ChatGPT account, with the same response times as for any other user. This management allows for a persistent and continuous conversation over a very long period of time, which makes all the difference! In theory, this approach also allows for a universal connection, adaptable to any AI with a web interface. Comparative tests are currently underway.

This personal experiment aims solely to explore new ways of using LLM-type AI. Integration with 3D worlds could be a future development, as the possible applications are countless. They rely on the immense knowledge base of these AIs and their ability to develop a rich narrative dialogue. Here are a few ideas: tours of museums, monuments, or landscapes, educational applications, Windows-like 3D environments, or simply discussions over a game of chess!

And you, what would you think of a 3D GPT?
What concrete uses come to mind?
Do you have any ideas for testing with GPT-5?
I can't post the video here, but you can find it by searching "Eidolon855" (without separating Eidolon from 855) on the web.

September 24, 2025 (Confirmation: I'm not good at chess !)

[tawdry vale]

it made a os

like a actual os

kind of

it can run micropython

and make basic ui

but it’s still very cool 🙂

[paper nest]

hi

[ruby scarab]

1