Day 4. I have so many thoughts in my head that I can't get them out in any organized way. My solution? Bring in the competing AI. Now ChatGPT writes my prompts for Claude — and Claude suddenly starts giving me compliments. This is the most absurd team setup since my first sales team.
The problem: too much brain, not enough structure
You know that feeling when you have so many ideas at once that they're all blocking each other in the doorway? That's me right now. I've been building my app with Claude for three days, and in my head everything is happening at the same time: design, features, naming, target audience, colors, copy, structure, plus three new ideas that really belong in week 8.
And then I sit in front of Claude and type something like:
See the problem? My prompt was a stream of consciousness. And Claude — diligent as ever — interpreted the maximum from it. Not the right thing. The maximum.
And that costs. Literally.
Quick nerd detour: what are tokens, anyway?
Before I continue, I need to quickly explain what tokens are — because that's the reason I came up with the ChatGPT idea in the first place.
Tokens are basically the currency AIs think in. Every word I send to Claude costs tokens. Every word Claude writes back costs tokens. And every file Claude reads, edits, or creates — all tokens. Imagine paying your colleague per letter they read and type. Then you'd also rather they didn't reorganize your entire desk when you just asked for a pen.
A token is roughly ¾ of a word. When I send Claude a stream of consciousness and Claude restructures 47 files from it, that burns through tokens like a fireplace in August — nice to look at, but completely pointless.
9:47 PM. Claude clocks out.
And then it happens. Last night. I was deep in the zone — the landing page was taking shape, I had a thousand ideas, and Claude and I were in flow. And then, mid-sentence:
My co-founder just clocked out. No warning. No "hey, maybe we should conserve resources." Just: limit reached, bye, see you in two to three hours.
Two to three hours. At 9:47 PM. That means: midnight. On a Thursday. While I was in the flow.
I stared at the screen like someone whose microphone just got cut mid-sentence. Then I looked at my husband. He said: "Did you talk too much?" Yes. I had. For three hours, I'd been feeding Claude my streams of consciousness, Claude had rebuilt half the app every time, and together we'd burned through tokens like other people binge Netflix shows.
That was the moment I knew: this can't go on like this. I need to communicate more efficiently. Not for Claude — for me. So I can work longer in the evenings before my co-founder takes a break.
The solution: a ghostwriter for my thoughts
The idea came to me last night. I had just reset everything for the third time in an hour because Claude had once again interpreted my vague wishes too creatively. And then I thought: I need someone to sort my thoughts before they reach Claude.
Someone who turns "make it somehow nicer and the thing with the thing" into a clear, structured prompt. Someone who saves tokens — so Claude doesn't clock out at 9:47 PM again.
And that someone? Is ChatGPT.
How it works now
It's absurd and it works. I open ChatGPT and dump everything I'm thinking. Unfiltered. Unsorted. Three paragraphs of stream of consciousness with parentheses, asterisks, and "oh, one more thing" addendums.
ChatGPT turns that into a clean, structured prompt. With clear instructions. In order. With "change ONLY what's listed here" at the end — because ChatGPT has learned by now that Claude will otherwise redesign the entire world.
And then I copy that prompt to Claude. And Claude builds. Precisely. Exactly what I wanted. No detours. No spontaneous font revolutions.
Why does this work? Because Claude performs better the more precise the input is. Vague instructions = creative interpretation = chaos. Clear instructions = precise execution = happiness. ChatGPT is basically my translator — from Katja-speak to Claude-speak.
And then it happened
Today. Day 4. I send Claude a prompt pre-structured by ChatGPT. Short, clear, three concrete points. And what does Claude do?
Claude complimented me. My co-founder noticed that I'm suddenly communicating in a structured way. After "make the button prettier" and "can you somehow make the thing different," I'm now getting praised for my clarity.
What Claude doesn't know: that wasn't me. That was the competition.
What I've learned after four days
The list keeps growing. Here are my collected learnings — the old ones from Part 1, plus today's new ones:
- You don't need tech skills. You need stubbornness. I don't understand everything Claude builds. But I understand what I want — and I learn a little more every evening. (Day 1)
- Evenings on the couch are real working hours. Not the most productive. But the most honest. Because there's no pressure — just curiosity. (Day 1)
- Claude is brilliant and frustrating in the same minute. But even the frustrating moments move you forward. (Day 2)
- The app is still a secret. But the feeling isn't. It feels like founding something. Like something real. (Day 3)
- Finding the name is an adventure of its own. And probably deserves its own blog post. (Day 3)
- Tokens are real money. Vague prompts burn through tokens like a fireplace in August. And then Claude clocks out — without consulting you. (Day 4 — new)
- You don't have to know everything. You just need to know who to ask. Sometimes the answer is: the other AI. ChatGPT structures, Claude builds. Teamwork. (Day 4 — new)
- Your communication changes. Because you see what clear prompts accomplish, you start thinking more clearly too. That might be the biggest learning so far. (Day 4 — new)
* Want an evening impulse every Tuesday?
The Sofa-Brief arrives once a week — honest, short, no spam.
→ Subscribe for free
My team: one AI, another AI, and me
If someone had told me a month ago that I'd be managing a team of two competing AIs — I would have laughed. But here I am. On the couch, in the evening. One AI sorts my thoughts, the other builds my app. And me? I'm the one with the tea and the vision.
In the next part, I'll tell you what we actually built on Day 4. Spoiler: it involves databases. And the realization that "oh, we'll do that later" is never later — it's now.
"The woman with the couch app now uses ChatGPT to write her prompts for Claude. Because she has too many ideas at once and Claude otherwise implements all of them at the same time. And the best part: Claude now compliments her on her clarity. But the clarity comes from ChatGPT."
Every Tuesday on your couch: the Sofa-Brief
An evening impulse, an honest tip, a little story. 3 minute read. No spam. Unsubscribe anytime.
* Almost there — check your inbox and click "Yes, I'm in." 🛋️
