Sahil | SimpleTechMan

I build apps with AI and show you exactly how.
🔧 Claude Code · Swift · React
AppStore:

[Sahil | SimpleTechMan @simple_tech_man]

I'm Sahil. I build apps with AI and document the whole process.

What I've shipped so far:
→ Steady (daily affirmations) — live on App Store
→ Joy To Do (task manager) — live on App Store

I share everything I learn here and on YouTube.
Follow along → https://t.co/8dHkio6FuM

[Sahil | SimpleTechMan @simple_tech_man]

Made a 4-minute video on this: where indigo-500 actually came from, which tells you can grep for, which "AI look" accusations are innocent (bento grids, dark mode), and the four lines to put in CLAUDE.md before any CSS exists.

https://t.co/MLvnnPD3my

[Sahil | SimpleTechMan @simple_tech_man]

Fix your purple problem and you land on cream, a tasteful serif and sage green. That's now the number one AI tell.

Chasing individual colours is a treadmill. The median moves every year. The centre doesn't.

Stop letting the model pick.

[Sahil | SimpleTechMan @simple_tech_man]

A model with no constraints doesn't give you its taste. It gives you the median of everything it read.

Purple isn't a preference. It's a centre of gravity.

That's why you get the same landing page from every model, they all read the same internet.

[Sahil | SimpleTechMan @simple_tech_man]

Nobody chose the purple gradient your AI keeps shipping.

Tailwind picked indigo-500 as a demo accent. Tutorials copied it. Starters copied them. All of it went to GitHub where the models learned what a website looks like.

Tailwind's creator apologised in 2025.

[Sahil | SimpleTechMan @simple_tech_man]

A skill is not fine-tuning.
Not memory.
Not magic.

It's a labeled instruction sheet the AI picks up when the task calls for it.

The steps you'd hand a smart new hire? That's a skill. Write them, add a one-line label, drop it in the folder.

[Sahil | SimpleTechMan @simple_tech_man]

An AI skill is just a folder with a text file inside.

No code. No retraining. No model surgery.

A folder and some plain instructions the AI reads at the right moment.

That's the whole thing.

[Sahil | SimpleTechMan @simple_tech_man]

The €0→€10K Claude Code screenshots are real. What they leave out: the other half the person did by hand.

The tool didn't make €10K. A person made €10K faster, because the boring half got automated.

[Sahil | SimpleTechMan @simple_tech_man]

Claude Code doesn't ship the product, pick the idea, or find the customers.

It writes code - brilliantly, fast, tirelessly. It's one worker on the team, not the whole company.

You're not replaced. You're amplified.

[Sahil | SimpleTechMan @simple_tech_man]

Picture a kitchen.

The AI is a great cook, it just doesn't have your recipes.

A skill is one recipe card.

You can hand it 500 cards and it stays just as fast, because it only reads the labels, then pulls the ONE card it needs.

[Sahil | SimpleTechMan @simple_tech_man]

You stop prompting. You start building the thing that prompts.

5-min breakdown: https://t.co/zpbFOwuN4R

[Sahil | SimpleTechMan @simple_tech_man]

The honest trap with AI loops: comprehension debt.

It runs faster than you can keep up, and one morning you don't know what your own project does.

The fix isn't "don't build loops."

It's: stay the engineer. Always keep a step where you check it too.

[Sahil | SimpleTechMan @simple_tech_man]

Claude Code is just this loop pointed at your codebase.

The tools usually arrive through an MCP server. Every lap spends tokens and fills the context window.

No magic brain. A familiar AI, given tools, put in a loop.

[Sahil | SimpleTechMan @simple_tech_man]

Broke down subagents the simple way: one desk, one room, one sticky note.

Wrote the cheat-sheet version with a copy-paste prompt → https://t.co/S3jwwlaee8

Prefer to watch? → https://t.co/K6qzZg76Ud

[Sahil | SimpleTechMan @simple_tech_man]

A subagent is a temp specialist you hire for one job.

It gets its own room.
Fresh, empty memory.
Does the messy work elsewhere.
Hands back one sticky note: the result.

Then it's gone. Your main AI's desk stays clean.

[Sahil | SimpleTechMan @simple_tech_man]

Attention is a budget.

The more you cram into one context window, the less sharp the answer on any single part.

A subagent is focus, on purpose.

One room. One job. One clean answer.

[Sahil | SimpleTechMan @simple_tech_man]

Everyone thinks AI gets dumber on big tasks because the model is weak.

It isn't.

It's a manager with one desk, and you piled 3 jobs on it.

The fix isn't a smarter model. It's a subagent.

[Sahil | SimpleTechMan @simple_tech_man]

The model never actually reads your words.

It chops them into tokens first — whole words, pieces of words, spaces, even commas.

"Understanding" starts only after your sentence has been broken into parts.

[Sahil | SimpleTechMan @simple_tech_man]

An AI agent is not a smarter brain.

It's a smarter setup around the brain. And that setup is a loop.

The model didn't change. What's around it did.

[Sahil | SimpleTechMan @simple_tech_man]

When you hit enter on a prompt, four things happen: tokenize, embed, attention, predict.

That's the whole machine. Everything else is detail.

It's not magic, and it's not Google.

[Sahil | SimpleTechMan @simple_tech_man]

Honest moment:

I almost didn't launch Steady because I thought "who am I to build a wellness app?"

Imposter syndrome is real.

Shipped it anyway. Now it's helping people start their day better.

Sometimes you just have to press publish.