Early access Working with a small set of design partners
Gonfire is an assessment platform built around the assumption that candidates will use AI — and that the way they steer the AI is the signal worth measuring. Real codebases, every interaction captured.
same task · two candidates live capture
Candidate A jamie@example.com
Candidate B priya@example.com
The problem
A few years ago a take-home told you something. Today every candidate ships a clean PR generated mostly by an AI assistant. The PR isn't the signal anymore.
01
Whether the candidate carefully designed the solution or one-shot prompted the AI to write it, the PR usually looks fine. By the time you find out which it was, they're three months in.
every PR looks AI-shaped
02
Inverting a binary tree never told you whether someone could design. Now that the actual job is "give the AI clear, well-structured context and review what comes back," whiteboards are even further from real work than they used to be.
Test invert a binary tree Job structure context for an LLM
03
How a candidate plans before prompting. Where they push back when the AI does the wrong thing. When they revert and try a different angle. None of this shows up in the final PR.
final PR prompts · reverts · reads · decisions where the signal lives
Before / after
Gonfire isn't trying to replace your entire interview loop — culture and team-fit interviews still belong on the calendar. We're trying to replace the part where you stare at a take-home PR and try to guess what the candidate actually did.
Today 0 Minutes per candidate · staring at a diff
→
With Gonfire 0 Minutes per candidate · reviewing the assessment
Today's take-home review
Send take-home prompt 5 min
Wait for submission 3–7 days
Read PR + reconstruct what happened 60–90 min
Schedule + run debrief about the take-home 60 min
Reviewer time~2 hrs
What you learnOutput looks AI-shaped. Unclear who steered.
With Gonfire
Send Gonfire link 2 min
Candidate works async, configurable
Review the assessment 15–30 min
Optional debrief, anchored to specific moments 30 min
Reviewer time~45 min
What you learnHow they planned, prompted, recovered, decided.
Attribution
We don't try to label individual lines as "human-written" vs. "AI-written." The candidate uses AI for the whole thing — that's the point. The label that matters is behavioral: which prompt led to which decision, where the candidate took the time to read the AI's output, and where they accepted it without review.
Behavioral Not "who typed what." Which prompt led to which decision, observed in order.
Decision points Where the candidate paused, reverted, or accepted AI output without review.
Per session The full interaction trail — not a single accuracy number to defend.
· [09:14] prompt: "explain the windowing logic" prompt
· [09:14] read AI output · paused 12s read
· [09:18] edited limiter.ts (+5 −2) edit
· [09:21] reverted edit, prompted again revert
· [09:24] prompt: "what edge cases am I missing?" prompt
· [09:27] accepted suggestion · no edits accept
· [09:31] ran tests · 14 passing test
vs. the take-home, as it exists today
| Today's take-home | Gonfire | |
|---|---|---|
| What candidates work on | Whatever you set up | A repo or zip you specify |
| AI tools used | Whatever the candidate uses, invisible to you | Captured end-to-end |
| What you review | Output PR | Full assessment + final submission |
| What you learn about process | Inferred from commit history, if you're lucky | Observed directly |
| Reviewer time | 2+ hours staring at a diff | ~30 minutes reviewing the assessment |
| Setup cost for you | High — env, prompt, rubric, evaluation criteria | Low — paste a repo URL, write a brief |
| Setup cost for candidate | High — local environment, dependency hell | Low — minimal setup required |
We don't position against HackerRank or CodeSignal — those platforms try to prevent AI use, which is the opposite of our approach.
Early signals
Quotes here as we hear from them. If you'd like to try Gonfire on a real role and tell us what's broken, reach out.
Pricing
Gonfire is in early access. We're calibrating the product with a small set of design partners before publishing pricing. If you want to use it on a real role, get in touch — we'll set up an account and walk you through it.
FAQ
The starter source can be any GitHub repo or zip. We've tested with TypeScript, Python, Rust, and Go.
Yes. The candidate setup page explicitly states that their AI interactions will be recorded and evaluated as part of the assessment. There's no covert capture.
Not yet. Custom rubrics are on the roadmap.
Reach out and we'll walk you through one. We're admin-provisioning accounts for design partners.
Get in touch
We're working with a few teams at a time. If you have a role open and a take-home you're tired of grading, we'd like to talk.
Early access · we respond within a day.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。