One real story a week, no repackaged news from the internet

Situation

Mrs. Sedláková runs a small accounting firm outside Ostrava, four people, one of whom handles new inquiries alongside the bookkeeping. We met this past spring for something else entirely, a consultation about their website, and over coffee she happened to show me an invoice for $430. "I don't even know if we actually need this, but I'm afraid to cancel it, what if something falls through the cracks afterward."

It was a paid add-on to their main accounting software, for tracking open client proposals. She'd signed a one-year term for a discount the salesperson offered at the time, and it had auto-renewed ever since, without anyone asking whether it still made sense. We sat down for an hour and went through what they actually used. In the end it came down to tracking response deadlines, an email reminder to the client, and an overview of where each proposal was stuck, three features out of the seven the package advertises on its own site. The rest, a competitor-comparison report, a Slack integration they'd never once turned on and I'm still not sure why it was even in there, kept running on its own, because it was part of the bundle and nobody had gone in to touch the settings.

What we did

I treated it from the start as a half-day fix, so I wouldn't overthink it. I asked the basic questions, who actually looks at that overview, how often, and what happens when a deadline slips. Her, once a week on Monday morning with her coffee, and when a deadline slips, they find out late and end up apologizing to the client.

They already had the foundation, a Google Sheet they kept updating by hand anyway, as a backup, because they never fully trusted the paid tool. I took that sheet as the single source of truth and built a simple page on top of it in Vercel, no login, just a link only they knew.

On top of that, two rules in Brevo, no complicated decision trees, just this:

I described this whole flow to Claude in one longer prompt, who sees what, when an email should go out, what the table should look like. I didn't write a single line by hand, I just kept refining it and testing it first on five old, already-closed proposals, to make sure it calculated deadlines correctly before connecting it to live data. The whole thing came together in one weekend, no new subscription needed, since they already had Sheets, Brevo and Vercel, or had them for free. Monday morning, the first time Mrs. Sedláková logged into the live version, she wrote back just this: "This is exactly it, thank you."

Where I messed up

The first version recognized won proposals by looking for the exact text "Won" in the status column. Except one of her colleagues had been writing "OK" instead, for a long time, and it never mattered while only a human was reading it. Thursday morning, over breakfast, Mrs. Sedláková forwarded me this:

At first I thought it was a one-off fluke. But in the automation log I found it had happened twice more that same week, nobody had noticed because those two clients simply hadn't replied. The automation was sending a follow-up reminder three months after signing, asking them to decide, except they'd already decided long ago.

I called Mrs. Sedláková that same day. I expected her to be annoyed, she was mostly tired, two years of dealing with small headaches like this from the old tool, and bracing for it to happen all over again with this one too. I told her exactly what had happened and why, and by evening it was fixed, status now gets picked from three fixed options, no more free text. Worse than fixing the code was figuring out who owed an apology.

I'd built the system around how the data should look in an ideal case. The reality is people type it in differently, in a rush, in shorthand, their own way, and that's exactly where the real mistake started. Now I ask first what people actually write into a field when nobody's watching, and only then decide what fields the table needs.

Try this

Open Claude and write exactly this:

When it answers, don't look for permission to cancel. Look for the moment it tells you something you didn't already know, like you're paying for three features but using one. That's exactly the moment you find out whether it's worth building yourself, or leaving as is.

"AI is an amplifier for a good foundation, not a rescue for chaos."

Karel