Did they actually do the work?

Written work stopped being proof. Anyone can submit something they didn't write and can't explain. Talk Me Through asks them about their own submission — out loud, for five minutes — and shows you whether the understanding is really there.

No candidate account. No download. Nothing recorded.

The signal you used to rely on is gone

A take-home task, a writing sample, a portfolio piece — these worked because producing them took the skill you were hiring for. That link is broken. You can no longer tell, from the artefact alone, whether the person in front of you could have made it.

You can't un-invent that. But you can ask them to talk you through it. Someone who did the work explains their own decisions easily. Someone who didn't runs out of road in about ninety seconds.

How it works

Three steps. The whole thing takes you under a minute and the candidate about five.

1

Paste their work

Code, an essay, a report, a design rationale — whatever they handed in. Questions are written from that exact submission, so they can't be answered from general knowledge.

2

Send them the link

It opens in their browser. No sign-up, no install, no scheduling. They speak their answers; they can type instead at any point.

3

Read the evidence

You get a per-question breakdown and the full transcript, so you can check the reasoning yourself rather than taking a score on trust.

What you actually get

Not a score out of ten. Evidence, with the transcript attached, so you can disagree with it.

Example report — illustration only
Mixed Confidence: medium · 6 questions · 4 min 51 s

Explained the caching layer and the retry logic fluently, including a trade-off that isn't visible in the code. Could not say why the database schema was denormalised the way it is, and described the error handling in terms that don't match what was submitted. Worth asking about those two areas directly.

Strong
Why retry twice before failing over? Gave the actual reason — the downstream timeout — without being led there.
Strong
What breaks if the cache is cold at peak? Described the exact failure mode and the mitigation they chose against it.
Partial
Why store the timestamps as integers? Right general idea, but attributed it to a reason the code contradicts.
Weak
Why is the orders table denormalised? Generic answer about performance; no reference to anything in the submission.

An illustration of the format, not a real candidate.

What it refuses to judge

A spoken check is a fairness risk if you build it carelessly. This one is built to look at one thing only: whether the substance of the answer matches the work.

It looks at

  • Whether the answer is specific to this submission
  • Whether the reasoning matches the decisions in the work
  • Whether they can handle a follow-up on the same point

It ignores

  • Accent, fluency, and first language
  • Hesitation, filler words, pace, confidence
  • Vocabulary, grammar, and anything hinting at background

Speech recognition makes mistakes, and the report is written to read through them charitably rather than punish them. The transcript is always included, so a result you doubt is a result you can check.

It never makes the hiring decision. It produces evidence for a human, and says so on every report. Audio is never recorded or stored — speech becomes text in the candidate's own browser, and only the words are kept.

Pricing

Free

While it's in early access. No card, no trial clock. It's new, and the honest thing is to let people use it and tell us whether the reports are actually worth paying for.

When that changes

Anyone using it during early access gets told well before pricing starts, and won't be charged for anything they ran while it was free. If you have a view on what this should cost, say so — that's genuinely useful right now.

Questions

Can't the candidate just use AI during the call?

They can try. The questions come from their own submission and ask about decisions and trade-offs rather than facts, and a follow-up lands immediately on anything vague. Reading a generated answer aloud tends to show up as fluent but unanchored — right-sounding, but never about the specific thing in front of them. It raises the effort a lot; it doesn't claim to be unbeatable, and the transcript is there so you can judge for yourself.

Is this a substitute for interviewing them?

No. It's a filter that tells you whether a written submission is worth trusting before you spend an hour of someone's time on it — and a prompt sheet for the areas worth probing in person.

What if the candidate can't speak, or would rather type?

Typing is available on every question, always, and never counted against them. The point is whether they can explain their work, not whether they can talk.

What do you do with the candidate's data?

No audio is recorded. Speech is turned into text by the candidate's own browser; only the text is stored, alongside the work you pasted, so the report can be shown. Nothing is sold, and nothing is used to train anything.

What kinds of work does it handle?

Anything you can paste as text — source code, essays, reports, research summaries, project write-ups, design rationales. If it arrived as a document you can read, it will work.