I built the thing I wished existed.

The story starts at a career fair, which is a bad place for stories to start. I was talking to a recruiter, friendly, mid-conversation, somehow we got onto how candidates were handling technical assessments. Everyone is cheating, she said, not really in a whisper. A couple students within earshot nodded. I nodded. We moved on.

I forgot about it for two weeks.

Then I was sitting in my own OA. Proctored tab, green-bar timer, honor-pledge checkbox at the top assuring everyone I wouldn't use AI. And I felt it: the urge. Not because I couldn't do the problem. Because everyone else was. Because the test was measuring something (can you solve this without the tool you're going to use every day for the next ten years) that didn't map to the job I was applying for. Because the people who were going to get the offer were the people who'd already decided the rule didn't matter.

I sat there with my hands on the keyboard and the test became a test of a different thing entirely. Not can you code. It was: are you going to be the chump?

That's the moment Promptster started. Not in a Notion doc, not in a hiring room. In an OA, hands hovering, realizing the question itself was broken, and that if it stayed broken, the people who got hired in my class were going to be selected for willingness to cheat, not skill.

The candidates aren't the problem. The test is. Companies are running an interview from 2018 on engineers who already work in 2026, then acting surprised when the signal collapses. Promptster is what falls out the other side of taking that seriously: stop pretending AI isn't in the room, instrument how the work actually gets done, and grade the process, not just the output everyone can now generate. It started as a hiring tool. A year of building it out across whole engineering teams taught me what the signal looks like over months instead of hours, and sent me back to the room where it decides the most.

What broke

For two decades, technical interviews worked because writing code was the bottleneck. If you could produce the code, we assumed you'd thought it through, because there was no other way to get there. Output was a faithful proxy for process.

That proxy is gone. Eighty-five percent of developers use AI coding tools daily. The candidate you hire on Monday opens Claude Code before they open Slack. Their job, your job, has become orchestration: model, tools, repo, judgment, all running in parallel, with the human pulling the threads together. The output is downstream of that orchestration. Grading the output now is like grading a chef by the plate while ignoring the kitchen.

Correctness is cheap now. Process is expensive.

Every legacy assessment vendor I've watched respond to this has done one of two things. They've bolted “AI features” onto a sandbox that doesn't look like the job, or they've banned AI from the interview entirely and called that rigor. Both are wrong in the same way: they treat AI as the problem instead of the context.

And the same proxy collapse happened inside the org, on people already on payroll. A manager who could once read the team's health off its output now looks at merged PRs a model half-wrote, dashboards that count tokens without explaining any of them, and a tool bill that doubled while nobody can say what it bought. The interview is where the signal broke first. It is nowhere near where it broke worst.

Why I built it the way I did

Promptster reads the sessions your engineers already have with their own AI agents (Claude Code, Codex, Cursor, the tools they use every day) and turns that history into the three numbers an engineering org already reports: cost, velocity and quality, on one timeline, with every config change and tool switch marked where it happened. Read-only, no proxy in the path, and your source code never reaches us. Because the work behind the first readout has already happened, the answer exists on day one instead of a quarter after install.

Three principles I refuse to compromise on:

Enablement, not surveillance. Individual numbers go to that engineer as private coaching and stop there; managers get team aggregates. A per-person ranking is a thing the product refuses to produce. Teams agree to install this because of that answer, not despite it.

Dev-to-dev, not HR-tech. I'm an engineer. I've been on both sides of the loop. If this product ever starts sounding like a vendor pitch (“AI-powered talent intelligence”), you have permission to email me about it. The vocabulary here is telemetry, signal-to-noise, orchestration, lead time. That's the job. That's the product.

Evidence over score. A score with no replay attached is a vibe with a number on it. We ship the receipts first; the score is the index, not the answer.

Where this is going

We're in design-partner mode through the rest of 2026. Twelve engineering teams, picked personally, on the record with me every week. Everything unlocked. Founding price locked through 2028. If we raise, you don't. I'm doing it this way because the measurement only gets sharper if the cohort is real, and the cohort is only real if I'm close enough to it to hear when we're wrong.

The hiring product this company started as is where it is headed again, and it is what the front page sells now. If you are interviewing senior engineers and the take-home stopped telling you anything, that is the one to look at. The team product runs on the same engine and lives at /teams: if you bought AI tools for your engineers and can't yet say what they changed, I'd like fifteen minutes to show you your own numbers. Either way, if it isn't a fit, I'll say so on the call.