From a Two-Year-Old Todo to a Live Product in One Day

In May 2024 I wrote a three-word Todoist note: “Wasm, Ads, Analytics” under the task Make online guitar tuner. It sat there for two years. Last night I said “let’s build,” and by the end of the session tuner.klokie.com was live, tested, and working well enough that tuning my own guitar with it genuinely surprised me.

The interesting part isn’t that AI wrote the code fast. It’s everything that happened before the code, and how much of my original three-word plan didn’t survive it.

The idea got interrogated before it got built

Instead of jumping to implementation, the session started with a YC-style office hours: who actually needs this, what do they do today, what’s the smallest thing worth shipping? Under questioning, my honest answers reshaped the product:

I’m not the user. I tune with a clip-on or by ear. The real user is a beginner holding a borrowed guitar who won’t install an app — and doesn’t know which string is which.
“Guitar tuner online” is a brutal market. Fender and GuitarTuna own the head term, and AI search summaries have eaten 15–25% of utility-page clicks since 2025. The winnable ground is long-tail tunings and shareable links, not the front page of Google.
Two of my three original words got cut. WebAssembly turned out to be engineering vanity — a well-chosen algorithm on the plain Web Audio API is plenty. And ads at launch would have damaged the one edge a newcomer has over the incumbents: a page that respects you. (Analytics survived, in cookieless form.)

Then the plan got attacked

Two different AI models reviewed the design adversarially before a line of code existed — and a second model, seeing only a summary, challenged the premises of the first. The best catches were things I’d never have found until users hit them:

Peg arrows can break strings. The original design showed “turn this peg clockwise” animations. But rotation direction depends on how the string is wound and which way you’re facing the headstock — confident wrong advice, aimed at the person least equipped to doubt it. The shipped version highlights which peg belongs to the string and says whether the pitch needs to go up or down. Strictly less flashy, strictly safer.
My success metric was wrong. “User tunes all six strings” counts a drop-D player retuning one string as a failure. The metric became per-string.
Adjacent guitar strings are five semitones apart — so an auto-detector that guesses within ±6 semitones can confidently point at the wrong string. The shipped snap window is ±2, with a manual selector as the escape hatch.

The code was the easy part

The build itself followed a boring, proven path: Astro static pages, the McLeod Pitch Method via an open-source library, one shared data model so that each new tuning is a data entry plus a page of copy. The part I’m happiest about is the test rig: Chromium can be launched with a fake microphone fed from a WAV file, so the entire beginner flow — tap start, grant mic, pluck a low E, watch the needle settle, get the checkmark — runs headlessly in CI with synthesized guitar tones. The classic failure mode of pitch detectors (reading a low E an octave high when a phone mic swallows the fundamental) has its own regression test.

A few hours after launch it grew from three tunings to seven — drop C, DADGAD, open G, whole step down — at a marginal cost of roughly one data entry and one page of copy each, because the review had forced that architecture in advance.

What I’d tell you to steal

Not the tuner (though please, tune your guitar). The pipeline: make the AI argue with your idea before it builds your idea. The building was one day. The two years of not-building were the expensive part — and the hour of adversarial questioning is what kept the one day from producing the wrong product.