An interview that looks
like the job.
No shared screen, no whiteboard, no timer counting down over an algorithm nobody writes any more. The candidate opens a link and lands in a working environment with an assistant beside them — and you get to watch how they use it.
24class Schedule:25 """A publish slot in the editor's local timezone."""26 slot: datetime27 tz: str2829 def next_run(self, tz):30 # callers pass naive datetimes31 base = self.slot.replace(tzinfo=None)32 if base < now(tz):33 base += timedelta(days=1)34 return base.astimezone(utc) # ← drops DST3536 def window(self, tz):37 return (self.next_run(tz), self.horizon)Four surfaces, all of them real
Nothing here is a simulation of a development environment. It is one, running in a container we built for this attempt.
The editor
The same editor that powers VS Code, with the syntax highlighting, multi-file tabs and keyboard shortcuts people already have in their fingers. Files save as they type.
The terminal
A real shell — several, if they want them. Install a package, run the test suite, start the dev server. Whatever the task needs, and whatever habits they bring.
The file tree
An existing codebase with more in it than the task touches. Finding the right file is part of the work, and so is deciding which files to leave alone.
Solara, in the panel
A conversational assistant, not a command line. It reads files, edits them and runs commands in the candidate's sandbox — and shows its work while it does.
Solara is a colleague, not a console
The candidate types in plain language: "a scheduled post fires an hour early after the clocks change — fix it." Solara reads the code, edits the files, and runs the suite. Every tool call it makes is a row in the transcript the candidate can expand and read.
The candidate can also ask for a plan first: in plan mode Solara reads the code and writes a plan, and changes nothing until the candidate approves it.
Which model it runs on is the candidate's choice, from a menu the problem author fixed when the problem was published — so every candidate on that problem gets the same options, and switching is a strategy decision rather than an advantage.
- Reads and writes files in the workspace
- Runs commands, and shows the output
- Asks before it runs a command; file edits apply at once
- Writes a plan first when asked, and waits for approval before it acts
- Can be interrupted mid-answer
- Keeps multiple conversations, so a dead end can be abandoned
- Runs on a metered budget the candidate can see as a bar, never a dollar figure
The change lands. What happens next is the signal.
When Solara edits a file, the edit is applied at once. It shows in the editor and in Source Control, against the commit the attempt started from. Solara asks only before it runs a command. What the candidate does with the change is theirs to decide.
They read the diff and agreed with it. Fast, but only as good as the reading.
They caught something and took it out — in Source Control, or by telling Solara why it was wrong. This is the single most informative moment in an attempt, and the old format has no way of producing it.
They took the shape and corrected the substance. Usually the strongest signal of all.
Our problems are written so that this matters. A well-designed brief has a place where the obvious answer is subtly wrong — a timezone that survives the tests, an index that looks fine until the table grows. A candidate who accepts everything sails past it. A candidate who is actually reading does not.
Set up in an afternoon
Pick or write a problem
Start from our library, clone one and make it yours, or author your own — brief, starter codebase, time limit, the dimensions you want scored and how much each counts.
Publish it, then invite
Publishing freezes a version, so every candidate sits the same interview even while you edit the next one. Invites are single-use links with an expiry.
Read the results
Scorecards land in the console as they finish. Invite the hiring manager, the team lead, anyone — seats are free and always will be.
Want to sit one yourself?
The fastest way to understand this is to be the candidate for twenty minutes. Ask and we'll send you an invite link.
request a walkthrough →Already comparing options? How five assessment platforms handle AI, including where each of them beats us. Or start with what AI fluency actually means.