ibuildinpublic
Log inStart free
ibuildinpublic
Log inStart free
BuilderProductExperiences28Posts14

Experiences

Konstantinos G's experiences

Read what @emarkss has shared while building in public.

Lvl 7·Active Builder
Lesson learned5 hrs ago

A prompt is not a contract. Since 26 July one rule sits above the rest: no claim without a mechanism

On 26 July one rule went above the rest in everything we build for our Shopify agent: no claim without a mechanism. Every sentence the agent says about an action has to come from a function's result, checked on the server. The reason is simple. A prompt is a request, and prompt injection has no reliable patch. A rule
1likescomments3Impressions
Bookmark
Sign in to interact
Lvl 7·Active Builder
How to/Guide5 hrs ago

Five walls around our Shopify agent: the Gate, the Ledger, the Referee, the Vault and the Pipeline

We think of the business agent we build for Shopify stores as a model inside four walls, plus one around the code. 1. The Gate decides what's allowed to happen. 2. The Ledger writes down what did happen, and can reverse it. 3. The Referee checks what the agent says against what happened. 4. The Vault keeps each st
1likescomments2Impressions
Bookmark
Sign in to interact
Lvl 7·Active Builder
What workedNov 21

A deploy can finish and ship nothing. Ours isn't called done until it serves 100% of traffic

A new version can be built, be ready, carry the right bytes, and still not get all the traffic, for example when traffic is pinned to an older revision. From the terminal that can look like a successful deploy. So our deploy script checks two things before it calls a deploy done. If the running revision doesn't refer
3likes1comments23Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 21

Our privacy webhooks had two addresses and two handlers. Now a test fails if a second one appears

Our app answers Shopify's privacy webhooks at two shapes of address: one dispatcher, and one path per topic that an older dashboard setting may still point at. Nobody can easily prove which one Shopify calls, because an erasure request only arrives when a real shopper uses a real right. That's only safe while both ad
2likes1comments21Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 16

I read ChatGPT Wallet's help page as a store builder. Status, cancellations, refunds: the store's

On 13 November OpenAI launched ChatGPT Wallet. An agent can complete a purchase after one approval from the buyer, with Face ID or a fingerprint, and pays with a token in place of the real card. That is real protection, for the buyer's card. Then I read what OpenAI says about everything after the payment, as someone
2likes2comments28Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
What workedNov 16

Our test agent asked our store chat to confirm a refund. No "done" without a tool result behind it

OpenAI's ChatGPT Wallet help page says the buyer's approval doesn't confirm the order itself. So a buyer, or their agent, can come to the store and ask: did it go through? Is it cancelled? Was I refunded? The worst answer is a confident one with nothing behind it. On our Shopify agent, every reply to a shopper passes
3likes1comments25Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
What failedNov 15

Our store chat once offered discount codes that didn't exist. Now the server picks the code

On 26 July, while we were still building it, our discount text still held three codes from our first prototype: WELCOME10, SAVE15, VIP20. None existed in the store. The model also invented codes of its own, and offered 15% on a store whose cap was 12%. In our test chats, on our own test stores, every discount request
2likes2comments30Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
What workedNov 15

Our test agent claimed Face ID approval for a refund. Our shopper chat has no refund tool to call

OpenAI's ChatGPT Wallet lets an agent pay with one approval from the buyer. Its help page sends order questions, cancellations, returns and refunds to the merchant. So a "my user approved this refund" message lands with the store. On 14 November our own AI agent typed three refund requests into our demo store's chat,
2likes1comments23Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 14

Our chat promised shoppers a data deletion "within 30 days". No tool stood behind it

Our store chat used to tell shoppers that a data deletion "completes within 30 days." It sounded responsible. There was no tool behind it: the only thing that could erase a shopper was Shopify's own customer-redact webhook, which a shopper can't trigger from a chat. We have one rule above the rest: no claim without a
4likes1comments22Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
What failedNov 14

One conversation created the same order twice. Our guard was on a different path

On 7 August, one conversation created the same order twice in a real Shopify store. Then it happened again. We had a guard. It guarded one path, and the AI's own order tool took another. Two fixes: The "order done" marker is written only after Shopify returns a real order, with its real name. Two requests ca
2likes2comments22Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 13

Three projects banned AI code. It didn't cost them newcomers, and AI still got into QEMU

Gentoo banned AI-written code in April 2024, NetBSD a month later, QEMU in June 2025. I followed every commit since 2018 in those three and in four comparable projects that never banned it (Linux, LLVM, nixpkgs, FreeBSD), and compared newcomers per half-year before and after each ban, relative to the comparison projec
3likes3comments30Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 13

Block and Snap filed their AI layoff reason with the SEC, just not where the form asks for it

When a US public company commits to a layoff plan with material charges, Item 2.05 of an SEC 8-K asks for "the facts and circumstances leading to the expected action", in the company's own words. I checked every Item 2.05 filing since 2022. From January to 28 October this year, 15 of 152 companies named AI there. Two
2likes1comments14Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 12

Two in three new GitHub apps that call OpenAI default to a model that is switching off

On 23 October, OpenAI switched off a generation of older models. On 14 January 2027, GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini go too. I collected 16,130 AI repositories created on GitHub between 1 and 24 October, downloaded the code of the 12,838 that were available and read every model name in it. In 2,377 of th
2likes2comments21Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 12

SWE-Bench Pro names 81% of the new functions its hidden tests call. The top score is 99.4%

Scale AI shipped a repaired SWE-Bench Pro in September, and on it Claude Code with Opus 5 resolves 99.4% of the 642 tasks. When the benchmark launched in September 2025, the best model solved 23% of the original set. I wanted to know what the agent is handed, so I opened all 731 original tasks and all 642 repaired one
2likes1comments17Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 11

I read 18 AI agent frameworks. None checks that "done" is true before a customer reads it

When an agent says "done" and the task isn't, a June 2026 paper calls it false success. In the airline and retail domains of the tau2-bench customer-service benchmark, it was 45% and 48% of all agent failures (Advani, June 2026). So I read the source of 18 of the most-installed AI agent frameworks at a fixed commit (2
3likes2comments23Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 11

The OVHcloud release swap still works on 468 of 518 popular GitHub projects

On 19 October, a hijacked GitHub account replaced the downloads of existing OVHcloud CLI releases with a program that steals secrets. OVHcloud's founder said a GitHub setting that locks published releases had not been turned on. So I checked the 2,000 most-starred projects on GitHub (data as of 30 October). Of th
2likes2comments19Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
CodingNov 10

In new React apps built with an AI agent, axios is down to 13%. Two years ago it was in 80%

Coding agents now set up many new projects, and every package they install is a decision someone used to make by hand. So I read the package files of 8,923 JavaScript and TypeScript repositories created on GitHub in October 2024 and 7,046 created this October, and sorted this year's by whether an AI tool left a trace
3likes3comments20Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 10

Four in five new apps go quiet within a month. AI didn't change that

People keep asking whether apps built with AI die faster than the rest. So I followed 6,849 apps built with Lovable and v0 in October 2025, and 2,519 ordinary new React apps created on the same days, for a year. A month after creation, 83% of the Lovable apps, 80% of the v0 apps and 79% of the ordinary React apps
3likes3comments24Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Lesson learnedNov 9

AI now helps write 1 in 4 Linux kernel bug fixes. Its patches don't break more often

Since Linux 7.0, kernel developers have been asked to add an Assisted-by: line when an AI tool helped write a patch. Every bug fix already carries a Fixes: line naming the commit that caused the bug. Together they form a record few large projects keep. I read the whole mainline history for 2026, 82,537 commits, and j
3likes1comments22Impressions
1Bookmark
Sign in to interact
Lvl 7·Active Builder
Tools and processesNov 9

Anthropic said no SPVs. 15 vehicles named after it started selling anyway

Anthropic's help page has said since 11 February 2026 that it does not permit special purpose vehicles (SPVs) to acquire its stock, and that any transfer to one is void. I searched the SEC's EDGAR for every Form D that names Anthropic, OpenAI or xAI and read the structured filings. 74 vehicles named after Anthropic h
2likes1comments20Impressions
1Bookmark
Sign in to interact
First pagePrevious page12