The long version: a whole session, what it costs, where your key goes, and what Hone can't judge.
How a session runs
A 40-minute senior interview on the same link shortener, in six steps. Scroll to move the clock. This one is written for this page.
You've pinned
100 reads for each new linknew
Redirect under 40 ms at P99new
Counts can lag a minutenew
You've pinned
100 reads for each new linknew
Redirect under 40 ms at P99new
Counts can lag a minutenew
3 min · scope
Ask before you draw
You ask how fresh the counts must be and how many reads come per new link. What you agree on gets pinned to the board.
You've pinned
100 reads for each new link
Redirect under 40 ms at P99
Counts can lag a minute
You
Can click counts lag by a minute?
Interviewerguiding
Yes, a minute is fine. Which number on the brief will shape your design most?
9 min · architect
Sketch the path
You draw the path a redirect takes. Your boxes are blue; the interviewer's circle is orange.
Interviewerprobing
That's about 2,000 redirects a second on average. What does each one cost this database?
14 min · architect
The next question follows the board
You add a cache, and the next question is about that box.
Interviewerprobing
What's the key, and when does an entry leave the cache?
21 min · defend
Defend your choices
You write each redirect also adds one to the link's click count. The interviewer pushes back.
You
Postgres can take it, and we can add read replicas.
Interviewerchallenging
So every redirect is also a write. What does that do to the database at peak?
30 min · scale
The interviewer changes a number
Half the users are now far from your servers. You put a CDN in front and work out the time.
Interviewerprobing
What does a redirect cost those users now, end to end?
Debrief · back to 21 min
The debrief rewinds to your biggest miss
Time ran out before failures came up. The debrief opens on your biggest miss, with the board as it was then.
Start here · biggest miss, turn 9
Every redirect also writes to the database, so the busiest path does work it doesn't need.
You wrote
“Postgres can take it, and we can add read replicas.”
A senior answer had
Counting stays off the redirect path: the API queues a click event and answers at once. A worker adds up counts in batches. Replicas spread reads, not writes.
Failure modes · not reached: the session ended first
3 min · scope
scope
architect
defend
scale
stress-test
stress-test · not reached
You've pinned
100 reads for each new linknew
Redirect under 40 ms at P99new
Counts can lag a minutenew
You've pinned
100 reads for each new linknew
Redirect under 40 ms at P99new
Counts can lag a minutenew
What you send, and when
A box you add stays blue, tagged “not sent yet”.
Your answer carries it to the interviewer, whose next question starts from what you changed.
Answer in words alone and the interviewer sketches what you describe, so you can move it, with a line like:
Interviewerguiding
I've sketched the queue you described; move it if I got it wrong.
Bring the question you were asked
Paste an interview question or a job description. Hone builds a private interview from it at the level you pick, with a hidden rubric so it stays fair. The text you paste isn't kept.
Your first build is free. After that it runs on your own key, for about $0.05.
You paste
Design ticket sales for a concert that sells out in two minutes.
2 sessions on Hone, with no account and no card, while today's free pool lasts (up to 6 a day from one network). It refills at 00:00 UTC (05:30 IST).
Then your own key
Add an Anthropic API key and a session costs about $0.20 (₹18), billed by Anthropic. Hone charges nothing today. An API key comes from the Anthropic Console and is billed apart from a Claude.ai subscription.
No subscription. Nothing for sale yet. About $0.20 from one measured session; longer sessions cost more.
Your key is kept in this browser's storage. Each request carries it to Hone's server, which uses it for that one call to Anthropic and never stores or logs it.
What you type and draw is saved for your debrief, where no other visitor can read it, and sent to Anthropic to run the interview.
What Hone can't judge
How you'd sound in the room. You type your answers; only your words and board count.
One company's bar. Grades follow a general mid, senior or staff bar.
The finest staff-level calls. A model reading a transcript misses things a panel would catch.
Whether you'd get the job. This is practice feedback, not a hiring decision.
For practice. Not for use in a live interview.
FAQ
Is the grading any good?+−
It's a practice signal, not a verdict. A model grades your transcript against a written rubric and the bar for your level. It only grades what you got to: anything you never reached is marked “not reached”, not guessed. Each point links to the turn it came from, so you can check it.
Do I need an account?+−
No. When you press start, your browser gets an anonymous visitor ID so Hone can keep your sessions. You may see a quick check from Cloudflare that you're a person. Add an email later to keep your sessions on another device.
Why might the free sessions run out today?+−
Hone pays for up to 20 free sessions a day across everyone, and up to 6 from one network, since an office or a mobile carrier can share one address. A free session also closes after 30 of your answers; on your own key there's no 30-answer limit. When the free sessions are gone, you can use your own key, or come back after 00:00 UTC (05:30 IST).
Does it work on a phone?+−
Yes. On a phone the interview and the board are two tabs. A laptop is easier, because you see both at once.
I'm early in my career. Which level should I pick?+−
Start with mid. It asks for sound parts and clear APIs. You can change the level any time.
Which model runs it?+−
Claude Sonnet 5.5, from Anthropic, for the interview and the grading. If it turns a request down on safety grounds, Anthropic's API can run that request again on another Claude model it picks.