---
title: "Three Times From Scratch"
description: "The third time I threw it away, I didn't even feel bad about it.
That's new."
url: https://www.mathys.to/writing/blog/three-times-from-scratch-24706/
section: blog
published: 2026-05-05
last_updated: 2026-05-05
---

# Three Times From Scratch

*The third time I threw it away, I didn't even feel bad about it.
That's new.*

### The PDF years

For a while, whenever Rob finished an initial conversation with a prospect, we'd make a PDF. Proposal-light. Professional-looking. It took real effort and it was never quite right.

Too much pricing, too soon. Too many assumptions dressed up as insight. A three-page explanation of what Playbook is, sent to someone who'd just spent an hour telling us what they needed.

We knew it wasn't working. We kept making them anyway, because what else do you do? You don't rebuild the whole thing because it feels slightly off.

Now I do.

![](https://d1ydx93l5afdxo.cloudfront.net/cms/250_b2982ba89218e5de_1280box.png)

### Version one

Rob has a method. After a prospect conversation he writes a short report — his read of the situation, the organization, the challenge. We wanted to use that report as the seed. Feed it to AI, extract the five signals that matter for this specific prospect, generate a minisite that actually speaks to them.

So I built an intake flow. Five steps. Thoughtful questions. Logic that held up on paper.

I watched it run once and knew immediately. It felt like an audit. Like the prospect was filling out a form to qualify for help they'd already asked for. I'd built something that interrogated people before it had earned the right to.

Threw it away.

![](https://d1ydx93l5afdxo.cloudfront.net/cms/250_2b6c964223f551db_1280box.png)

### Version two

Flipped it. Rob uploads the report, AI extracts the signals, draft appears instantly. No intake. Just: here's what we think we know about you.

The copy came out confident. Too confident. Whole paragraphs that spoke with certainty about things we'd only inferred from meeting notes.

Rob read it and said: *this sounds like a cold email from someone who thinks they know me.*

That's the sentence that killed version two. Back to zero.

![](https://d1ydx93l5afdxo.cloudfront.net/cms/250_e219041dca597e2d_1280box.png)

### Version three

This one does something I didn't plan for.

After Rob uploads his report, the AI extracts the five signals — but before it generates anything, it shows him which signals it found easily and which ones it had to guess at. Where the notes were thin. What was missing.

*Next time, ask about decision-making structure earlier. Budget context was hard to infer.*

Rob is getting better at his intake conversations because of a feedback loop I didn't design on purpose. It just appeared when I built the thing and looked at what it was doing.

That's the version we're using.

![](https://d1ydx93l5afdxo.cloudfront.net/cms/250_5a780d320df879df_1280box.png)

### The neutral voice in the room

Something I didn't expect: how useful it is to ask AI to critique what you just built.

Not to help you build it. To tell you honestly if it works.

I've been building things for twenty-five years. When a colleague looks at a prototype and says it isn't working, there's a layer of defensiveness you have to push through first. Your ego is in the room. You're managing the relationship alongside the feedback.

With AI that layer is just... thinner. Ask the right questions — *does this feel like an assumption? would a prospect feel understood or interrogated?* — and you get a straight answer with nothing social attached to it.

The willingness to hear feedback clearly is what actually drives iteration. Removing the friction from critique is underrated.

![](https://d1ydx93l5afdxo.cloudfront.net/cms/250_2096bb5038ed2c6e_1280box.png)

### What made it hard

The signal-extraction piece is genuinely difficult. Figuring out what counts as a meaningful behavioral signal versus background noise, in a short report written quickly after a meeting — that's not a solved problem.

I used to love hitting walls like this for competitive reasons. If it's hard for us, it's hard for anyone who tries to copy us. Now I love it for a different reason: hard problems are where the real learning is. You can't shortcut your way to an insight that only shows up through building.

![](https://d1ydx93l5afdxo.cloudfront.net/cms/250_b41849605d85c386_1280box.png)

Three full rebuilds. In the old model — developer, sprint, delivery — one of those would have been a crisis. A planning failure. Something to explain.

Here it was a Tuesday. Two of them, maybe.

That's the shift. Not that you can build faster. It's that you can afford to be wrong, follow the insight, and throw away the thing you just made without it costing you the project.

The PDF never taught us anything. The rebuilds taught us everything.

Mathys is co-founder of [Playbook](https://beplaybook.com) and builds tools at the intersection of behavior change, voice, and reflection.