Training for the SF Marathon with the app I'm building

In twelve days I run the San Francisco Marathon. For the last twenty-one weeks, every mile of the build-up has lived inside Navara — which is either the best product research method I know or an elaborate way to make my own bugs personal. It has been both.

A plan that watches you back

The plan started in March: base, build, peak, taper across 21 weeks, topping out around 40-mile weeks. The feature that mattered most in practice is the one I now can't train without — demonstrated-load recalibration. Navara doesn't hold you to what a template guessed in March; it watches what you actually run. Show up above target two weeks in a row and the plan ratifies the higher load instead of letting the following weeks sandbag you. Nineteen of my twenty-one weeks landed on-track, and the two that didn't were absorbed instead of cascading — the plan bent where older training plans would have snapped.

The taper argument

Last week I connected Claude to my plan and asked it to review the taper. That turned into a whole story of its own — the short version is that my taper this week is one an AI coach designed and I approved, tap by tap, in the app: volume down to ~30 miles, but with a short tempo and a race-week sharpener (3×1 mile at goal pace) keeping the legs honest. Two easy runs, a recovery jog, a 12-mile long run with the final stretch drifting toward goal pace. It is, honestly, a better taper than I would have written for myself.

The number I'm not refreshing

Navara's race predictor is built on critical-speed maths over your real best efforts — not a lookup table from one time trial. It has watched the long runs get longer, faster, and lower-heart-rate through June, and it currently has opinions about July 27 that I am choosing not to look at more than once a day. Ask me again after race day.

Dogfooding, with teeth

Training for a real race inside your own beta is a forcing function nothing else matches. This month alone it surfaced a sync bug that had quietly frozen my plan's cloud copy, a duplicate-detection flaw in the Strava import that only a fourteen-year archive could expose, and a taper philosophy the research disagreed with. Every one of those is fixed for the beta because I hit them first, with a race on the line.

The SF Marathon is July 27. The beta opens around the same time — if you want to train for your race the way I trained for mine, early access is below. I'll write the race report either way; that's the deal with building in public.