Humza 🇵🇰 🇨🇦

I like to build things. e/acc to infinity and beyond

[Humza 🇵🇰 🇨🇦 @0xHumza]

The art of prototyping clickable mobile apps with claude.

There's a lot of project-level instructions and files, including a design system, to achieve these results on plain text prompts.

Thread on instructions coming soon.

[Humza 🇵🇰 🇨🇦 @0xHumza]

https://t.co/cJG8WdFJVm

[Peter Gorski @peter_gorski]

Founders: this appears to be a scam, or at best a form of blunt idea mining.

We completed their application a few months ago—a fairly lengthy and detailed forand never received any follow-up. Just radio silence. Multiple follow-ups from our side also went unanswered.

[Humza 🇵🇰 🇨🇦 @0xHumza]

never deleting this app

[Humza 🇵🇰 🇨🇦 @0xHumza]

@aarondfrancis You can always ask Claude for a good loop for someone who wants to get looped into running loops.

[Humza 🇵🇰 🇨🇦 @0xHumza]

Tip: don't make the model design your layout from scratch, it eats tokens.

Pull a few site screenshots you like, feed them to ChatGPT, and ask for an ASCII layout diagram first. Cheapest way to settle what goes where.

Then have it render a hi-res mockup and run 5 iterations in your chosen design language. You get to the final faster for fewer tokens.

And by now you have a visual reference, design system, and detailed instructions, so now you can feed it into any model for implementation. The brainstorming is done, ez pz

[Humza 🇵🇰 🇨🇦 @0xHumza]

@Semah____ Sweet, just wanted to share my experience so you don't waste as many tokens as I did.

[Humza 🇵🇰 🇨🇦 @0xHumza]

Yes, I know I don't need so many sub-agents with 1M context windows. They are there for emotional support.

[Humza 🇵🇰 🇨🇦 @0xHumza]

The good:
Claude workflows with Opus 4.8 are great at synthesizing and figuring out "what is the next step?" and then solving it end to end.

The bad:
10% weekly limit down in one session.

The ugly:
I unfortunately don't work at Anthropic, therefore I will never be able to fully experience the potential of this beast in its true, glorious form.

[Humza 🇵🇰 🇨🇦 @0xHumza]

@rezoundous As the IPO nears, whenever that might be, the limits will be gradually decreasing to maximize shareholder value,

[Humza 🇵🇰 🇨🇦 @0xHumza]

@krispuckett I mean we can debate the recommendations, but it's kind of sad to outsource your thinking to a machine, especially when it comes to forming sentences.

AI tone is too visible in this post.

[Humza 🇵🇰 🇨🇦 @0xHumza]

@jp54362 If you're still manually looping coding agents in 2026, you're basically the middle manager workflow.

The move is building loops that bully the loops into bullying agents into prompting themselves.

[Humza 🇵🇰 🇨🇦 @0xHumza]

Well, I think that's the fastest way to run into your limits on any paid subscription, regardless of the provider.

And realistically speaking, I think most of the engineers don't actually have such a workflow that requires them to have agents working around the clock to deliver on things.

Human-in-the-loop is an unsolved problem because I think it's not a problem for a lot of people at all. Some loops do require a human to be present in them.

[Humza 🇵🇰 🇨🇦 @0xHumza]

@TheJerzWay Not that straightforward. Significant ties is a thing.

Regardless, it’s a beautiful place to be at and honestly nothing short of a privilege to call it home :)

[Humza 🇵🇰 🇨🇦 @0xHumza]

most likely context bloat + and incorrectly using models & thinking mode for tasks.

You need a local routing policy to use sub-agents with the right models + thinking mode when working with it.

e.g complex code edits = GPT-5.5 medium/high thinking mode

draft a PR = GPT 5.5-mini medium

etc.

Copy/paste this post, paste it into codex and ask it to analyze sessions for the last week and suggest a sub-agent routing policy which is configured globally on your machine.

you are welcome :)

[Humza 🇵🇰 🇨🇦 @0xHumza]

When I started using AI initially, I would spend anywhere between 20 to 25 minutes writing prompts, and it used to be in cursor. For most of the time, it took me a fraction of that time (< 5 mins) to do those tasks.

Being able to speak my thoughts into reality is definitely a game changer.

203 wpm 🫡

[Humza 🇵🇰 🇨🇦 @0xHumza]

Last month in action.

[Humza 🇵🇰 🇨🇦 @0xHumza]

@DanielSmidstrup Having 11 paying users and losing one of them.

[Humza 🇵🇰 🇨🇦 @0xHumza]

@rezoundous Codex limits are actually slightly worse than Claude now, and it's kind of sad to see. I remember there was a time I burned like 2.2 billion tokens during a 5-hour Codex session, and my weekly limit barely moved.

[Humza 🇵🇰 🇨🇦 @0xHumza]

Working on something with my co-founder, and honestly, we haven't even looked at a single accelerator yet.

We have a bunch of customer demos lined up. There's a high chance we will be able to convert some of them into paying customers. If so, that's the highest validation.

Everything else comes second. Nothing validates your product-market fit more than a bunch of people willing to give you money to solve their problems for them.

Saves you so much time from filling random applications.

[Humza 🇵🇰 🇨🇦 @0xHumza]

@shadcn The art of shipping composable software for the masses to use, never an easy feat.