Open Training

Pickleball sessions a group of advanced players (around 4.0 and up) can run without a coach, one phone page per session: https://open-training-site.vercel.app/

A free, non-commercial side project, built in my spare time by Volker Max Meyer, a product manager and service designer in Berlin, with Claude Code.

The problem

I coach the advanced group in my pickleball club. I play along on one court, and a playing co-coach runs the second. Good drills are spread across hundreds of YouTube videos, most of them written for two players, and none of them is a plan a group can pull up on court. My assumption is that players who want to train without a coach mostly end up playing matches instead. The test under Status is there to check it.

What it does

The start page asks what the group wants to get better at and offers nine goals, such as "Win the dink battle". Each goal opens a 30-minute session of three drills. A drill card shows one sentence, the part of a coach's video that demonstrates the drill (it plays inside the card after a tap), the rules, how to make it easier or harder, and what good looks like. A short form at the end asks whether the session worked without anyone explaining it.

Where it comes from

This site is one of three interfaces on a private training system I built for my group.

                  ┌──────────────── private ────────────────┐
 YouTube videos   │ Base: skill ladders, drills, weekly     │──▶ Coach: plan a session, plus
 + research on    │ themes, session frame, tactical and     │    a card for the co-coach
 teaching tactics │ mental cues, strategy playbook          │──▶ Open Training (this repo)
                  │                                         │──▶ Overview: what the base
                  │ Quality loop: guard scripts, feedback   │    contains, on one page
                  │ after every session                     │
                  └─────────────────────────────────────────┘

In November 2026 the base held 120 drills distilled from 205 reviewed videos, all in one shared vocabulary. The interfaces only read from it, so a new one is cheap to add, and the base was built with more of them in mind. It started as a Google Doc for my own training and grew into a system once I saw that other players could use the same material. The product view of the whole system is in docs/prd.md, a snapshot from 6 November 2026.

How I keep AI output honest

Claude Code wrote most of the code and much of the content. I set the goal, the rules and the priorities, and I approve or reject the work at fixed points. The checks themselves run by machine:

• 154 tests, among them a contrast test that fails below 7:1 and a test that keeps DESIGN.md and the CSS tokens in sync. They run before every publish commit and every push; red tests stop the release.
• A leak guard on every commit and push, so private notes and paths never reach this public repository (see below).
• A version number and a changelog line for every push, and a script that waits until Vercel serves the pushed commit.
• In the private base, three scripts check the vocabulary, the cue consistency and the links on every change. Every new drill's video link is checked against frames from the video before the drill is saved, because an audit in July 2026 found 26 of 100 links pointing to the wrong moment.

Product decisions

• Cards follow the video, which is usually a pair drill, and groups of four get one warning line: never two balls with pace on one court. Translating every drill for four players had been the most expensive step, and it produced forms nobody had played.
• The video plays inside the card and stops where the drill ends, because court time is short.
• Taps only: no typing, no login, no cookies. Text and background keep at least 7:1 contrast so the page stays readable in the sun.
• The look is playful on purpose. The players should enjoy opening it, and a sober dashboard style was tried and dropped.
• How much a card should show is being tested with a version that has sliders. Testers set it up the way they want to read it between two drills and send a screenshot.
• The AI kept proposing test programmes sized for a startup. I cut the first test down to one question: is it needed?

Status

Live since November 2026 with nine sessions. Groups beyond my own have not used it yet. A lean test through the club's coaches runs until January 2027: do groups play a session and come back to it?

How the site works

private notes ──publish script──▶ src/content/sessions/*.json ──git push──▶ Vercel build ──▶ static site
                 (validate, extract,            (leak guard runs on commit and push)
                  allowlist fields)

• Astro, static output, no adapter. Tailwind for styling. No external fonts or third-party scripts; a YouTube player (youtube-nocookie.com) loads only after a tap on a video.
• Content is JSON, not copied notes. scripts/publish.ts reads the sessions marked for publishing and builds each one from named fields only, so private notes and links never reach the repo. The site validates the JSON against src/lib/session-schema.ts.
• Feedback goes to a Tally form; the link carries the session slug as a hidden field. Answers stay in Tally.
• Visitor statistics come from Vercel Web Analytics: no cookies, no personal profiles. Loading times come from Vercel Speed Insights (since 2026-11-09), also cookieless; both are named on the privacy page.
• Every deploy is checkable. The site serves version.txt with the commit it was built from; scripts/check-deploy.ts polls it until the pushed commit is live.

Publishing

npm run publish:site -- --dry-run   # show what would change
npm run publish:site                # validate, write JSON, guard, tests, commit, push
npx tsx scripts/check-deploy.ts     # wait until Vercel serves the new commit

This is the beginning of the README. Read the rest at the original. Read the full README ↗