Seven prompts. Take all of them.

Hey — and welcome to Sunday Signal.

Most prompts give you an answer. You read it, you use it, it's gone.

A few do something better. Some hand you a file you'll paste into every conversation for the next year. Others walk you through a thing you'd never have started alone.

I used seven of those this week. They're all below, in full, no gate.

1. A billion people now use the same assistant
On August 11 Google announced the Gemini app crossed one billion monthly users, the fastest-growing product in the company's history. The usage detail is the interesting part: 63% of people talk to it out loud, and one in five live sessions goes beyond voice into a camera feed or a shared screen.

Why it matters for you: A billion people have access to roughly what you have access to. Whatever separates your output from theirs, it is not the tool.

2. Meta gave away a model that runs on your laptop
On August 10 Meta released Muse Glimmer, a 30-billion-parameter model under an Apache 2.0 licence, built for agent work and small enough to run on a single consumer GPU. No cloud, no subscription, no per-token bill.

Why it matters for you: Access is not the moat either. Serious capability is now free and local for anyone who wants it.

3. Two days after the billion-user news, Google shipped another model
Gemini 3.7 Flash landed on August 13, three weeks after the previous version, at half the price of its predecessor through the end of the year.

Why it matters for you: The models change every few weeks. A prompt you wrote in March still works in August. Whatever you build on top of these tools should be the part that survives them.

Two kinds of prompt are worth keeping. Everything else is a party trick.

The first kind hands you a file. You answer some questions, and what comes back is a paragraph you paste at the top of every chat from now on. The work compounds instead of evaporating.

The second kind gets you through a door you'd have stood outside of. Not because the prompt is clever, but because it turns one terrifying task into four ordinary ones. Most people never build the thing they've been describing for two years, and it isn't a skill problem. Nobody broke it into steps small enough to start.

The prompts that go viral are almost never either of these. They produce one impressive paragraph and then nothing.

I put a game on the App Store this week. I have never written a line of code.

It's called Spire. Start to finish it took about four hours: I described what I wanted in plain English, an AI designed it, another wrote every line, and I tested it on my phone and reported what broke. Then it broke again right after launch, and I had the AI help me fix that too.

I'm not telling you that to impress you. Four hours and a handful of tools is not an achievement, it's a demonstration. The interesting part is that the whole thing is four prompts repeated with different nouns, and they're printed below.

Elsewhere this week: three carousels went out, each one a single copy-paste prompt tied to something real. And I killed my own cold email system after nineteen straight weekdays where it wrote emails and not one of them got sent. The copy was fine. The step where a human presses send was the thing that never happened.

Seven prompts. If you only do one thing today, do prompt 1 and actually answer its questions. Everything else compounds off that.

These are the ones I promised in yesterday's post. The whole sequence, in order.

1. Describe it, and make it ask you things

"I want to build [describe your idea in plain English]. Before you write anything, interview me until you genuinely understand it: what it does, who it's for, every screen, and what should happen when something goes wrong. Ask me one question at a time. When you're done, write it up as a build plan."

The one-question-at-a-time part matters. Ask for everything at once and you get a wall you won't read.

2. Make it plan before it builds

"Here's the plan we agreed. Before writing any code, list the files you'll create and what each one does, in plain English. Wait for my go-ahead before building anything."

This is the step people skip. It's also the step that stops you receiving two thousand lines of something you didn't want.

3. Report the bug like a witness, not a victim

"Here's exactly what I did: [steps]. Here's what I expected: [x]. Here's what actually happened: [y]. Don't guess at a fix. Tell me what you think is wrong and how you'd confirm it, before you change anything."

"It's broken" gets you a guess. This gets you a diagnosis. The don't-guess line is doing most of the work.

4. Make it walk you out the door

"I want to publish this to [the App Store, or the web]. Walk me through it one screen at a time. Assume I have never done this before. Stop after each step and wait until I tell you it's done."

That's how the game got submitted. I did not know what I was doing at any point, and it did not matter.

The tools, since people always ask: Claude for the design, Claude Code for the building, RevenueCat for payments. Building a website instead of an app, swap the last two for Vercel and Stripe.

I'd rather you hear this from me than find it out at midnight.

You do have to know what to ask for. The AI will build exactly what you describe, including the parts you described badly. You will fix things you don't fully understand, and you'll have to make peace with that. Mine broke right after launch, in front of real people. Four hours is the build, not the learning.

None of that is a reason to skip it. It's a reason to start with something small enough that breaking it doesn't matter.

5. The memory file — make it interview you

The problem: you re-explain who you are at the start of every conversation.

"Build my memory file. Interview me: who I am, what I'm building, how I like to work, and 5 things you should never do. Format it as a note I'll paste at the start of future chats."

Answer honestly. What comes back looks something like this:

"Kalpesh runs a one-person studio building AI products. Currently shipping a mobile app and a web tool. Prefers plain English, tables over paragraphs, and wants to be told when something is wrong rather than agreed with. Never: invent numbers, pad an answer to look thorough, or claim something works without checking."

Six lines. Paste that at the top of your next chat and the first answer changes. One rule that makes it compound: write down the WHY on every decision, not just what you decided.

6. The voice profile — stop sounding like a press release

The problem: default AI writing opens with "thrilled to announce." Nobody talks like that.

Find three things you've written that sound like you. Paste them in, then:

"Here are 3 samples of my writing. Study my voice: sentence length, rhythm, the words I use and avoid, how casual I am. Write a one-paragraph 'voice profile' I can paste before every request. Then wait for my next task and write it in that voice."

Save the profile. Paste it into any chat where you want to sound like yourself.

I wanted this on my phone, so it became an app: https://apps.apple.com/us/app/caption-coach-ai/id6765720293

7. One idea into a week — stop posting things once

The problem: you make something good, post it once, and it's gone by tomorrow.

Paste your idea, or a transcript of you talking about it, then:

"Here's one idea. Break it into a week of content: 1 short-form video hook + script, 1 carousel outline, 3 short text posts for X, and 1 longer story post. Keep my angle, don't water it down, and make each one able to stand alone."

I wanted this done automatically from a podcast, so it became a tool: https://getamplifi.live/

Reply and tell me which one you ran. If you build something with the first four, I want to see it more than anything else in my inbox.

Fourteen Saturdays in. The arc so far: ship, constrain, build, compound, patch, subtract, ask, remember, price, staff, equip, doubt, show, and now keep. The models change every three weeks. A prompt you wrote in March still works, and the file it produced still works too.

That's the quiet thing I've learned doing this in public. Almost nothing I've built is clever. I wrote a few things down once instead of re-explaining them fifty times, and I broke the scary jobs into steps small enough to start. The compounding did the rest.

The long-form is nearly there. One piece of filming left. No date from me, since I've missed a few.

If any of these seven were useful, forward this to one person who has been describing the same idea for two years without starting it. The first four are for them.

Quietly, deliberately. One Saturday at a time.

— Kalpesh
Founder, Sun Ventures

P.S. Work with me. Everything above came out of building real things for real use.