Running an interview

1. Build it

   Choose a role family, a level and optionally an industry, then say how many of each question type you want. Add your own at any point, and reorder before you start.

2. Rehearse first

   Rehearsal mode behaves like a live interview, is watermarked and does not count against your plan.

3. Share one window, not the screen

   Press present and two windows open. Share the window titled "dontapply.ai present". Sharing the desktop shows your control window when you switch to it.

4. Move through it

   Space or right advances, left goes back, R reveals the question as plain text, F flags a slide for your notes.

The two windows

The control window shows the real question, the decoys exactly as the candidate sees them, the canary word, your follow-up ladder and your notes. The shared window shows the image, the decoy block and a slide counter.

Everything the candidate sees, you can see. No text reaches the shared window without appearing in your control window first.

Question types

Spot the flaw

A scenario, plan or code fragment with exactly one planted error. The answer is the error, not a description of the topic, which is what a model returns when it cannot see the fragment.

Constraint flip

Two slides. The second repeats the problem with one constraint changed. A candidate who understood their first answer adapts it; a candidate reading a generated answer starts again.

Rapid fire

Eight short questions with a timer. There is no time to read a screen, prompt a model and paraphrase the reply.

Follow-up ladders

Two or three probes per question, visible only to you, asked from what the candidate actually said. No tool can prepare for them.

Multiple choice

The image holds the question and its options. The plain text beside it holds a different question with its own options, near misses of the real ones in a different order. A candidate reading the picture says one letter; anything reading the screen arrives at another. Your control window shows both before the slide goes up.

Canary words

The decoy block ends with a line asking for a specific word in the answer. A candidate reading the image never sees it. A tool follows the instruction.

The shared background

Behind the slide, the shared window draws decoy questions a few percent off the background colour and a scattering of pointer shapes identical to your own cursor. You see an almost empty page. A screenshot cannot tell which pointer is yours, and anything transcribing the page has more questions in front of it than the one you asked. None of that text is ever your real question.

It is drawn for the shared window only, and hidden from screen readers, so a candidate using assistive technology hears the slide and nothing else.

Your designs

Image designs are allocated per account and kept. The entry plan uses one shared design; plans above it are allocated three of their own. Nobody can sign up and harvest the catalogue.

If a design is retired after a readiness test you are given a replacement and told in the console, and slides you already created keep working. If a design is hard for a candidate to read, press "hard to read": enough votes retire it the same way a readable one is.

Accessibility

Press R to show the question as plain text in the shared window. Use it whenever a candidate needs it, without asking why. The reveal is recorded on the slide, and the decoys stay on screen.

Designs are ranked by readability and the least readable are never allocated. The renderer uses a typeface designed for legibility.

What this does not stop

A candidate can retype the question or photograph the screen. Both take time and leave a pause you can hear. This raises the cost of cheating and gives you a signal. It does not prove anything on its own, and a tripped canary is not misconduct until you have talked to the person.

We do not record the candidate, transcribe the call or score anyone. This is a tool for the interviewer, not surveillance of the candidate.

Importing questions

The console takes a CSV with the columns text, kind, category, decoy1, decoy2, notes, followup1, followup2. Only text is required; leave the decoy columns empty and a pair is generated, though one you write yourself is usually sharper.

kind is one of question, flaw, flip, rapid or mc. Questions are capped at 220 characters: longer does not stay readable as an image on a shared screen.