Skip to content

The interview where candidates actually build with AI.

Watch how candidates scope, decide, and orchestrate inside a real AI coding session, the signal a traditional technical assessment can't show.

Book a 15-min intake See a real session
The problem · as of 2026

The interview no longer measuresthe job.Two candidates ship code that passes the same tests. One pasted what the model gave them; the other actually reasoned about the problem. Traditional technical assessments can’t tell them apart.

A
Candidate ASenior · 6 yrs
00:42
  1. 00:05prompt“write a stripe webhook handler in node, return 200 on success”
  2. 00:11ai pasteroutes/webhook.ts · +47 −0 · pasted verbatim
  3. 00:42submitsubmitted: no retry test, no duplicate-event test
Promptster signalPrompt-proxy
B
Candidate BSenior · 6 yrs
06:20
  1. 00:09readscanned spec: at-least-once delivery, signed payloads
  2. 00:38planidempotency key: event.id · DB unique constraint
  3. 01:05prompt“verify signature first, then idempotency-key lookup”
  4. 01:48editroutes/webhook.ts · +52 −0
  5. 02:30editmigrations/…add_webhook_events_unique.sql · +8 −0
  6. 03:12testreplayed same event twice: second is no-op ✓
  7. 03:55prompt“what if Stripe retries before our DB commit lands?”
  8. 04:30editwrap in transaction · row-level lock on event id
  9. 05:10testconcurrent replay (race): one side-effect only ✓
  10. 05:48revertdropped early-200 return · breaks at-least-once
  11. 06:20submitsubmitted with notes on retries + reconciliation
Promptster signalArchitect
opportunity cost · estimated

One bad senior hire isn’t a misstep. It’s a quarter.

  1. 6+ moLead time burned per bad senior hireRoadmap slip from a hire who can't actually orchestrate.
  2. ~$200KFully-loaded cost of one mis-hireComp, onboarding, ramp, severance, opportunity. Estimated.
  3. 47 hrsEngineering time per closed-loop interviewPanel × prep × debrief, summed across the team.

Every false negative costs another loop. Every false positive costs the team’s lead time. Today’s interview can’t tell which is which.

How it works · 3 steps · ≈ 90 seconds of setup

Candidates work in their real editor. You get the whole transcript.

Built with your team

Before a single candidate gets a key, we build the assessment with you: your stack, your seniority bar, the bugs your team actually ships. We design it together to surface the signal that matters: how a candidate orchestrates an AI agent on the kind of work you'd really hand them, not a LeetCode puzzle they can prompt their way through. You give us the role and a few representative problems; we draft it, you sign off. Design partners get it included.

01

Your team emails the candidate a key.

You email the candidate one assessment key. No account, no portal. They install the CLI and type the key. Telemetry starts on their consent, never before.

# candidate · one-line install (macOS · Linux · Windows)
❯ promptster start PST-8R2M-4KVN
02

They work normally.

Claude Code or Codex, in the editor they already use: VS Code, Cursor, Neovim. Same prompts, same tool calls, same flow. A lightweight Go binary routes their agent through our proxy and registers native hooks. Zero impact on their workflow.

● telemetry · active · proxy + hooks
prompts · responses · tool calls · file diffs
× keystrokes · × clipboard · × webcam · × screen
03

You open the case file in your browser.

A server-side transcript with per-change attribution, a tiered read on process and outcome, and a copy-paste-ready brief your panel reads in two minutes instead of running another loop. Every claim is replayable.

# reviewer · in your browser
app.promptster.ai / sessions / PST-8R2M-4KVN
→ solid operator · 1 dimension flagged · chain verified ✓
Insight, not surveillance

Everything a reviewer needs. Nothing a candidate should dread.

No proctoring, no lockdown, no keystrokes or webcam, and candidates consent before anything is captured, in the editor they already use.

Defensibility · the question legal asks first

A score that decides who you reject
is a legal instrument.

The moment a number influences who gets cut, you're operating an automated employment decision tool. NYC's Local Law 144 and the EEOC's disparate-impact guidance both apply, and both turn on one question: can you explain the decision? A proprietary rubric and a percentile can't. "The model said 61" is the answer that loses.

We built Promptster so the answer is evidence, not a shrug. The rubric is open-source: a base you can read and fork, with dimensions and weights we tune to what your team rewards. Every factor links to the exact replay moment that moved it, and the score never rejects anyone on its own: a human acts on it, which is the posture these laws are written to protect. When legal or an auditor comes asking, you hand them the transcript.

Read the rubric on GitHub

Under the hood it's a classifier trained on how senior engineers orchestrate an AI agent: scoping before writing, naming tradeoffs, knowing when not to delegate. The output is a per-dimension read tuned to your funnel, your role and your bar, not a global leaderboard, and it only ever sees the work: prompts, tool calls, decisions, never a résumé, a name, or demographics.

Orchestration classifier · v0.4PST-3K9X-7FQM
Solidoperatorprocess read
Task framingStrong
ComprehensionStrong
Direction qualityAdequate
SteeringAdequate
Verification loopDeveloping
Fix integrityAdequate
Context managementStrong
Ecosystem leverageAdequate
One dimension flagged: verification loop. Replay 44:55. Candidate soft-delegated to the agent.
ai_fluency_v1 · tiered, not scored · calibration ongoing with design partners
the score · what it is, what it isn't
What it is
  • Decision support for a human
  • An open-source rubric, tuned to your bar
  • Every dimension → a replay moment
  • Tiers with stated evidence, not a score
What it never is
  • An automated reject
  • A black-box "trust us" score
  • Cross-role rankings
  • Résumé, name, or demographic inputs
We're not a compliance product, and we won't pretend a v0.4 model has passed a bias audit. We're built so you, or an independent auditor, can run one, because every decision is already evidence.
Why we're not HackerRank

A different category, not a
prettier leaderboard.

The comparison below is on what each platform actually measures, not which dashboard ships the slickest charts. HackerRank and CodeSignal grade the finished code. We watch how the engineer built it.

Comparison of HackerRank, CodeSignal, and Promptster across six dimensions of technical hiring assessment.
DimensionHackerRankLegacy assessmentCodeSignalLegacy assessmentPromptsterProcess telemetry
AI-tool postureDetect-and-block. Cat-and-mouse with copy-paste.Detect-and-block. Lockdown + plagiarism checks.Embrace. Measure how the candidate orchestrates the AI on purpose.
What you actually getA final diff and a pass/fail score.A final diff plus a percentile on an IQ-style scale.Process telemetry: prompts, tool calls, attribution, decision points.
Candidate environmentSandboxed editor in a locked-down browser.Sandboxed editor in a locked-down browser.Their own editor, their own repo, the AI agent they already use. The job, basically.
Evidence behind the decisionThe score. Trust it.The score and a percentile. Trust them.Replayable session, line-level attribution, server-side record.
Auditability of the scoreBlack-box rubric, proprietary.Black-box rubric, proprietary.Open rubric. Every rationale links back to a moment in the replay.
Skills it actually measuresAlgorithm trivia and timed problem-solving.Algorithms plus IQ-style cognitive proxies.Orchestration, judgment, AI-tool fluency: the work they'd actually do.
Deep divePromptster vs CodeSignal vs HackerRank vs Codility
AI-tool posture
HackerRank
Detect-and-block. Cat-and-mouse with copy-paste.
CodeSignal
Detect-and-block. Lockdown + plagiarism checks.
Promptster
Embrace. Measure how the candidate orchestrates the AI on purpose.
What you actually get
HackerRank
A final diff and a pass/fail score.
CodeSignal
A final diff plus a percentile on an IQ-style scale.
Promptster
Process telemetry: prompts, tool calls, attribution, decision points.
Candidate environment
HackerRank
Sandboxed editor in a locked-down browser.
CodeSignal
Sandboxed editor in a locked-down browser.
Promptster
Their own editor, their own repo, the AI agent they already use. The job, basically.
Evidence behind the decision
HackerRank
The score. Trust it.
CodeSignal
The score and a percentile. Trust them.
Promptster
Replayable session, line-level attribution, server-side record.
Auditability of the score
HackerRank
Black-box rubric, proprietary.
CodeSignal
Black-box rubric, proprietary.
Promptster
Open rubric. Every rationale links back to a moment in the replay.
Skills it actually measures
HackerRank
Algorithm trivia and timed problem-solving.
CodeSignal
Algorithms plus IQ-style cognitive proxies.
Promptster
Orchestration, judgment, AI-tool fluency: the work they'd actually do.
Design partner program · by invitation

Twelve teams get
founding spots.

Paarth, founder
Promptster · building it with you

I'm a recent UT Austin grad. Last year I applied to 20+ software engineering jobs, almost all opening with an OA or take-home, and no matter how much I studied, people around me kept outscoring me. When I asked how they prepped, the answer was: “don't study, just cheat.” I started asking recruiters their side, and nearly all said the same: since the AI boom, they can't screen candidates properly anymore.

So I stopped applying and built Promptster. The final code shows who can ship; the session shows who's exceptional. If you hire 5+ engineers a year and your screen can't tell paste from craft, I'd love to build this with you: twelve teams on the founding cohort, small enough that I'm in the room for every one.

Book a 15-min intake and we'll talk through your roles. If it isn't a fit, I'll say so on the call.

  • Assessments custom-built to your taste. We source and tailor problems to your stack, your seniority bar, and the bugs you actually ship. You don't write the questions; we do, and you sign off.
  • Weekly founder session. 45 min, on the record, about your roles and what the transcript is actually telling you.
  • You shape the roadmap. The features you ask for get built, under your team's name on the changelog.
founding price · lockedbilled monthly
Contact us
Founding rate · held below list
Cancel anytime. They don't
price lock
Your founding rate holds through 2028, even when we raise.
Book a 15-min intake
15-min intake · first assessment free
FAQ · the objections worth answering

Honest answers,
not marketing answers.

  • What if my engineers use Cursor, Codex, or Claude Code?
    Claude Code and Codex are both fully supported and capture full process telemetry: prompts, tool calls, file diffs. The signal you grade is the same across them, so split-agent teams are no problem. Cursor is where the boundary sits, and it falls where the model key does: assessments run on a key you supply, down one metered path, billed to you. Cursor sends its Agent and Edit traffic to its own backend, where that key is never used and never metered — so we support Cursor as an editor and instrument Claude Code or Codex inside it. Candidates keep their own editor and their own setup either way.
  • What does the candidate experience, and what if they refuse?
    Their normal editor, their repo, their flow: no browser lockdown, no proctoring. The session opens with an explicit consent screen listing exactly what we capture (prompts, tool calls, file diffs) and what we don't (keystrokes, screen, clipboard, webcam). If they decline, no session runs and nothing is captured; you see the decline in your dashboard and can offer an alternative.
  • Do I have to write my own assessment problems?
    No. We custom-build assessments to your stack, your seniority bar, and the bugs your team actually ships: you give us the role and a few representative problems, we draft it, and you sign off. Design partners get this included.
  • Can we change what the rubric rewards?
    Yes. Every team starts on our open-source base rubric: the dimensions and anchors live in a public repo you can read and fork, and it stays current as the agents evolve. If your bar is different, you care most about how a candidate steers a model that has gone wrong, you want test-first discipline to dominate the verification loop, or your stack rewards specific tools, we tune the dimensions and anchors to match how your team actually works. The base tracks the frontier; your overlay reflects your standards.
  • How long does setup take?
    About 15 minutes on the intake call. We generate an assessment key, you email it to candidates, and they install the CLI with one command; sessions stream back to your dashboard. No ATS integration required.
  • How do I compare reads across different roles?
    Per-role classifiers on a shared calibration model, so a read is always within-role, never cross-role: "strong on verification loop" means strong against the backend bar, not against every candidate you have ever run. Every dimension is tiered rather than scored, and each tier is visible with the session moments behind it, so you can audit how each role is being read instead of comparing two numbers that were never on the same scale.
  • What happens after the design-partner period ends?
    Your founding rate stays locked through 2028, and you keep access to everything from the design-partner period: no feature claw-back. If we raise list prices, you don't.
Further reading
On the record · signed · replayable

Read the process,
not just the commit.

Twelve founding teams will ship this with us. A technical screen that can't tell paste from craft isn't neutral. It's a ~$200K coin-flip you won't catch for months. If you hire 5+ engineers a year, we should talk.

Book a 15-min intake See how it works