Scope
Clarify the task: who uses the system, what it must do and what it need not.
An AI interviewer takes you through seven stages while you draw the design. Every question, answer, hint and diagram is kept, so your next attempt starts from what the last one showed.
Free for individuals · Runs through your own AI assistant
You estimated 10,000 writes per second. What happens to the write path when the cache is unavailable?
An example of a question. The assessment is AI practice, not a hiring decision.
The interview
The interviewer moves on when a stage is done, and notes what you agreed, what you missed and where a hint was needed.
Clarify the task: who uses the system, what it must do and what it need not.
Agree on load and volumes: requests, data size, growth. The numbers stay on record.
Name the operations the system offers, what they take and what they return.
Choose the storage and draw the tables or collections, the keys and the queries they serve.
Draw the services, queues and caches, and how a request passes through them.
Go into the hardest part: a bottleneck, a failure, consistency under load.
Answer for the trade-offs: what breaks first and what you would change with more time.
Four levels: junior, middle, senior and staff. You choose the level, and the result is given for it.
How to start
The interview runs in the chat of your AI assistant; the diagram and the record live in Keystone.
Create a practice project and connect Claude Code, Codex or another assistant that supports MCP. Keystone shows the settings to copy.
Send one command with a topic and a level, like “online shop, middle”. Without a topic the interviewer offers a choice.
When the interview ends you get scores for seven criteria with evidence, a result for the level and three improvements linked to your diagram.
Why not a whiteboard
A mock interview on a call ends when the call does. Here the whole session stays, and two attempts can be put side by side.
| Mock interview on a call | Keystone | |
|---|---|---|
| After the session | The board is wiped; notes, if somebody took them | Every question, answer, hint and weak spot is kept |
| Feedback | A general impression, from memory | Seven criteria scored from 1 to 4, each with evidence from your answers |
| The design | A photo of the board, at best | A diagram snapshot at the end of every stage |
| The second attempt | Starts from scratch | Compared with the first: criteria, hints and time per stage |
| When | When a colleague or a paid interviewer has time | Any time, at the level you choose |
A person who has run real interviews still notices what an AI does not. The practice prepares you for that conversation; it does not replace it.
Your own assistant: Claude Code, Codex or another client that supports MCP. You connect it to a Keystone project, and it interviews you in its chat. Keystone does not include or sell a model.
No. It asks, pushes back and gives a hint when you are stuck; hints are recorded. It does not change your diagram unless you ask.
No. It is an AI practice assessment for the level you chose: a way to see weak spots before a real interview, not an employer’s verdict.
Yes: name any topic, like a URL shortener, a chat or a payment system, and one of four levels. Without a topic the interviewer offers several.
The practice is part of the free plan for individuals. The AI assistant runs on your own subscription or key.
The interview text, the scorecard and the diagram snapshots of your practice project. A practice project needs no source code, and Keystone never receives any.
Start with one task
Pick a topic, choose a level and see where your design holds and where it does not.
Start practising