Cali

Wow if Apple gets access to Mythos, we might finally get a working search bar. Jokes aside, I think a ton of software is about to get pwnd and re-written from scratch. x.com/AnthropicAI/st…

@boristane Very relevant read btw anthropic.com/engineering/ha…

I'm building a rather big feature I'm using ai for everything instead of just building the feature I'm working on an extensible primitive to make it more versatile ngl ai sucks at this type of work the scope is wide and requires long term thinking, and despite writing and reviewing pages and pages of markdown files it still: - doesn't fully complete tasks (despite explicit todo lists I ask it to cross check) - ignores existing patters in the codebase (although explicitely spelled out in the plan document), and - takes every shortcut it can my rule of thumb now is, if an agent works over 45mins on a task, it's probably gonna be ass

@boristane Are you using GPT 5.4 xhigh? Unlike Opus 4.6 I feel like it doesn’t cut corners and has more depth. And do each of your plans have multiple phases?

I’ve been showing everyone how seamless and quick payments can be with @WalletConnect Pay using our latest SDK It just works… Scan -> Approve -> Done! My favorite Wallet to demo has been @moneda_com who did a great job with their iOS release supporting it natively 👌

Just tried @WalletConnect Pay for the first time. 🤯 Scan, pay, and that’s it. No card. No crazy process. This is the future of finance. Another epic WalletCon for the books.

@theo And being a successful founder is harder

Technically speaking, being a founder is easier than getting a job

AI-generated code tends to produce massive PRs so we deliberately fight that with small, phased PRs. Each one gets real attention, better tests, deeper review. A 2000-line LGTM ships bugs but ten 200-line PRs ship quality.

The real risk of AI isn't that it replaces you. It's that it makes you atrophy. You stop thinking through decisions because the agent did it for you. Then someone asks "why?" and you've got nothing.

I write about AI-assisted engineering, the path from mid-level to staff, and building a life that sustains high performance. Join the newsletter if you're interested in this kind of content calikurt.com/newsletter

If you take one thing from this: write the brief before you prompt. 3 sentences. What you want, what the constraints are, and what "done" looks like. Try it on one task this week and see what happens when you give AI your best input instead of your laziest. Full breakdown in my latest newsletter (link in reply).

Before anyone else sees the code, I review my own work with Claude Code. I ask it why something was implemented a certain way. I ask it to review from an SRE perspective. I ask about test coverage gaps. I have a back-and-forth discussion and push improvements before the team review even starts.

Then I break the work into phases. Each PR is small and digestible. AI-generated code tends to produce massive PRs, so we've deliberately worked against that. The agent has clear context for each phase instead of juggling everything at once. Small PRs get real attention, better tests, and higher quality.

Most developers write a vague prompt in Claude Code,, skim the output, and ship it. I did the same thing early on, and it cost me. A subquery that nearly exhausted our database. A technical doc I couldn't defend in front of my own team. Here's what I do differently now. 🧵

After the design doc, I write an implementation plan as a markdown file and commit it to the repo. This plan has code snippets, database migrations, the exact order PRs will be created, and a testing strategy. My first PR was the plan itself. My peers reviewed it, and we caught issues before a single line of implementation code existed.

Before I write a single prompt, I write a design doc. Last week I built a merchant management API. I spent an hour exploring two architectural options with Claude Code, challenging assumptions and poking holes in trade-offs. Then I shared the doc with my team and we aligned before any code was written.

@Capetlevrai Si tu te retrouves a utiliser des prompts de manière récurrente, converti les en skills ou commande avec Claude Code. Sinon check Obsidian

Vous utilisez un outil spécifique pour stocker vos Prompts genre un prompt manager ou prompt builder ? Je n'en ai pas mais je sens que ce serait bien d'avoir une app dédié à ça que je peux utiliser rapidos, comme un logiciel de prise de note mais dédié à ça

I put a lot of heart into my technical writing, I hope it's useful to you all. 📌 Here's a pinned thread of everything I've written. (much of this will be posted on the Claude blog soon as well)

We are proud to be a part of @Mastercard’s partner program and build the future of stablecoin and crypto payments together.