# My Running Coach Is a Claude Skill

> A 10k training plan that lives in Notion, reads my Apple Health runs, and updates itself: run, sync, re-predict, adjust, all driven by one sentence.

Published: 2026-06-04
Canonical: https://tomasparizek.com/2026/06/04/self-updating-running-coach.html

I wanted to run a 10k in under 40 minutes. These days you'd just grab an app for that: Runna or Garmin Connect will build you a plan and coach you through it. Instead I had Claude Code build me one from scratch, ground it in my own running data, and wire it up so it keeps updating itself as I train.

What I ended up with is less a training plan than a small automation system: a couple of Claude skills that read my health data, design the plan and write it into Notion, then keep it current as I train. The running goal is just what I pointed it at. The interesting part is how the pieces fit together.

## Built From My Own Data

The first thing the system does is ground the plan in real data, not a template.

Claude pulled my last nine weeks of activity from Apple Health and analyzed 20 actual runs. The data access runs through my [HealthExport CLI](https://github.com/TParizek/healthexport_cli), wired up as a skill, which I [wrote about here](/2026/04/05/making-agents-friendly-service.html). It found an easy pace of 5:30–6:00/km at a heart rate of 130–143, a genuine aerobic base left over from my marathon years, plus a couple of harder May efforts around 4:50–4:55/km. It also noticed my benchmark 10k was run on a hilly forest trail with 200 m of climb, so my flat-road fitness was probably closer to 44–46 minutes, a bit better than the 47 I'd quoted.

That detail mattered. The plan got calibrated to my physiology and my habits, not to an average runner's.

## It Asked Before Building

Before writing a single row, the agent did two things I'd want from any coach. It researched current best practice for a sub-40 10k: periodization through base, build, peak, and taper, an easy-to-hard split of roughly 75/25, and the standard menu of quality sessions (threshold, VO2max intervals, race-pace work). Then it interviewed me.

The questions were the high-leverage ones, not a generic intake form: how many days a week I'd actually run, how long my long run could get, and how I wanted to schedule. It folded the answers in, laid out the whole plan for approval, and only then built anything.

## The Plan

What it produced is 17 weeks long, moving through Base, Build, Peak, and Taper, with about 75% of running easy. Two quality sessions a week, one easy or cross-training day, and one or two strength sessions that taper off near race day. No fixed weekdays: I asked it to track completion per week, so the plan bends to whatever my week looks like.

| Phase | Weeks | Focus |
|---|---|---|
| Base | 1–5 | Aerobic foundation, early tempo work |
| Build | 6–11 | VO2max intervals, longer threshold reps |
| Peak | 12–15 | Race-pace work, sharpening |
| Taper | 16–17 | Reduce volume, stay sharp, race |

The paces are anchored to the goal and recalibrated from my data:

| Zone | Pace (min/km) | Use |
|---|---|---|
| Easy / recovery | 5:40–6:10 | easy & cross day |
| Long run | 5:15–5:45 | weekend base |
| Threshold | 4:25 → 4:10 | tempo |
| 10k goal pace | 4:00 | race-specific reps |
| VO2max | 3:50–3:58 | 1 km reps |

Two calibration races are built in: an all-out baseline 10k in week 2 to anchor the starting prediction, and a tune-up 5k in week 13 to recalibrate near peak. Both feed the prediction model directly.

## How It Lives in Notion

Everything sits in a single Notion page. The top of it shows the goal, the race details, the weekly template, and one prominent line: my current predicted race time.

Below that are two databases. A Training Log with one row per planned workout (51 rows, 17 weeks times three sessions) that tracks status, actual distance, pace, and heart rate. And a Prediction Log, a time series of every predicted finish time with the reasoning behind it, so I can watch the trend of my projected race time across the whole block instead of just the latest guess.

## The Feedback Loop

This is my favorite part. I had Claude build a reusable skill called `sync-runs`, and now when I say "sync my runs," it runs the whole loop:

- Pulls my runs from Apple Health since the last sync.
- Classifies each one by intensity from pace, heart-rate zones, and distance.
- Matches each run to that week's planned slot and updates the row: Done, Modified, or Missed, plus actual numbers and a one-line note ("10 s/km faster than target at lower heart rate, fitness trending up").
- Recomputes my predicted race time from real data, preferring my most recent all-out effort projected to 10k via the Riegel formula (a standard race-time predictor), then logs it.
- Proposes adjustments to upcoming weeks, nudging paces faster if I'm ahead or easing them if I'm struggling, and shows me the changes before writing anything.

![The self-updating training loop: run, sync, measure, re-predict, adjust, with a dashed arrow looping back from adjust to run.](/public/images/blog/2026-06-04-self-updating-running-coach-feedback-loop.svg)

A one-off plan became a closed loop I drive with a single sentence: run, sync, measure, re-predict, adjust. Most training plans are dead the moment you deviate from them. This one absorbs the deviation and tells me what it means.

## Takeaways

- **Build the loop, not just the artifact.** A static plan is dead the moment you deviate. A system that re-syncs, re-predicts, and adjusts stays useful as real life happens.
- **Compose small skills into a system.** One skill reads my health data, another reasons over it and writes to Notion. The whole thing is assembled from reusable parts.
- **Ground it in real data, not a template.** The plan came from my own runs, paces, and heart rate. That is the difference between generic and personal.