Andy Ke

My own productv0.2 preview, live demo

Pocket Tennis Team: a personal tennis coaching team on Claude

It runs on the player's computer today; phone capture and access are next, hence "Pocket". No real players have used it yet.

Players get advice from lessons, YouTube, teammates and parents, and keep almost none of it. I designed and built a Head Coach and six specialists, seven AI team members, each a Claude skill (a packaged set of instructions Claude follows): you talk after practice, the team remembers, and you get a small plan where every item says why.

Coach reply time

About 3 min to about 20 s

The time until the coach replies, measured in my own end-to-end tests with the live AI. Bookkeeping still runs in the background afterwards.

The team

7 members

Seven team members, each a Claude skill with its own job and rules, behind one conversation with a Head Coach.

Voice

36 voices

Natural on-device voices, English and Chinese, so the coach's replies are spoken back.

A pro has a team. A club player has none.

The problem

Advice from everywhere, a plan from nowhere

Club and junior players hear from lessons, YouTube, teammates and parents. Almost none of it is remembered, connected or turned into a plan. A pro has a coach, an analyst, a fitness coach and someone tracking every match. A club player has none of that.

Sam, 44, adult player

"If I have to type it, I won't."

Sam plays for fun and will never open a spreadsheet. So I made it voice first: Sam talks after practice or a match, and the team turns that into a training diary and a small plan.

Mei, 46, parent of Alex, 14

"Tell me what to say in the car, and what not to."

Mei runs it for Alex, a competitive junior. She isn't the player, so she gets entry deadlines, load and body checks, and help with the talk after a match.

The rules every screen follows, and what each one costs

A coach, not a dashboard

Conversation is the main surface, not charts. Each reply is signed by the team member who wrote it, so the player knows who decided what.

Every plan item has a Why

Each focus links to the player's own words, the member who decided it and a success target. A plan the player can't trace doesn't earn trust.

Small, honest load

At most three tennis focuses and one fitness block a week. That means saying no to bigger plans that would look more impressive.

One good session isn't proof

It takes repeated evidence to level up, so the plan changes slower than a chatbot would.

Health first

Tightness means monitor, pain pauses the related work, limping means stop now. Juniors get growth guardrails. Health guidance is labelled "not medical advice".

Private by default

Plain files in the player's own folder, and the app runs on their computer. Messages go to Claude for answers, with no cloud server or account of mine in between, even though that makes phone access harder.

Adults run it

Claude accounts are for adults, so a parent runs it for a junior. I cut the teen role until a policy review on minors.

Seven team members, each a Claude skill

I split a fuzzy job, coaching between lessons, into members with clear ownership of files and decisions. Replies are signed by the member who wrote them, and the player's real coach outranks videos and general tips. Player Scout, Fitness, Nutrition and Stats join later.

Team memberWhat they doJoins when
Head CoachThe front door. Listens to anything, answers fast, puts health first and calls in the right teammate.Every message
Team ManagerInterviews the player and sets up the team folder and the starter plan.First setup
Journal KeeperTurns messy notes into a clean training diary and spots repeating patterns.After practice or a match
Coach AnalystRemembers what the player's real coach and favorite channels teach, and checks the player's ideas against it.When a lesson or video comes in
Progress CoordinatorEach week decides what to keep, level up, simplify, rest or swap.Weekly review
Strategy CoachOwns the final goal and the long-term roadmap to get there.When the goal or roadmap changes
Tournament CoachHandles the event calendar, entry deadlines, game plans, debriefs and reflections.Around tournaments

How it works

The app page talks to a local server on the player's computer, which runs Claude with the plugin's skills. The player's folder is the memory, and the app renders from one file, state.json. My first real test took minutes per message, so I split each message into two passes: the coach's reply time went from about 3 minutes to about 20 seconds in my own end-to-end tests with the live AI, and the bookkeeping now runs after the reply.

One message, two passes

playerThe player talksVoice or text after practice or a match. The browser's speech service turns voice into words.
ClaudeFast replyA read-only Head Coach pass reads the plan and recent messages, then replies with the health rules built in.
appSpoken backThe app shows the reply and speaks it in one of 36 natural on-device voices, English and Chinese.
ClaudeBookkeepingA second pass runs the skills in the background so the journal, plan and calendar match what the player was told.
filesSaved as filesMarkdown, CSV and JSON in the player's own folder. The app refreshes from state.json when bookkeeping is done.

A Claude plugin with seven skills and an auto-open hook, a small standard-library Python server, one HTML page, and Kokoro, an open-source voice model that runs on the device.

How I designed it, and what the tests changed

  1. Personas and jobs

    I wrote Sam and Mei and mapped their top tasks, from telling the coach how practice went to knowing what to say in the car.

  2. Clickable prototypes, v1 to v3

    I reviewed each version as product owner before wiring in the real AI, so the flow was settled before the model was.

  3. Persona testing

    AI agents role-played Sam and Mei through onboarding, daily training and tournaments over five rounds, while a designer agent fixed what they found.

  4. Real end-to-end tests

    Running the live AI changed the design: the two-pass reply, the natural voice and several safety fixes all came from these tests.

Smoothness scores out of 5 from AI agents role-playing each persona, not real players: 2 to 3 in the first test, 4 to 4.5 in the final round of five. Both persona agents' final verdict was "ready to show". Real players come next, in the friends beta.

PersonaFirst test (v2)Final test (v3)
Sam (adult player)onboarding 3, daily 2, tournaments 2interview 4, plan review 4.5, home and daily 4.5, tournaments 4.5
Mei (parent)onboarding 3, daily 3, tournaments 2interview 4, plan review 4.5, family home 4.5, tournaments and health 4.5

Screens from the live demo

Live demo, sample data, in a desktop browser.

Interview with a live plan board: As the player answers, the plan builds on screen right away, without waiting on the AI.
Interview with a live plan board As the player answers, the plan builds on screen right away, without waiting on the AI.
Plan review: Every item has a Why? that shows the player's own words, who decided it and the success target.
Plan review Every item has a Why? that shows the player's own words, who decided it and the success target.
Home: Today's one thing, the next steps, and a big button to talk to your coach.
Home Today's one thing, the next steps, and a big button to talk to your coach.
Plan: The final goal, the long-term roadmap from the Strategy Coach, and the current training block.
Plan The final goal, the long-term roadmap from the Strategy Coach, and the current training block.
Team: Each teammate with a Ready status, "Set up & train" and Ask, so the player can go to one directly.
Team Each teammate with a Ready status, "Set up & train" and Ask, so the player can go to one directly.

Where it stands