LLenny's Podcast
← All frameworks
InnovationDan Shipper

Build Your Own Senior-Engineer Benchmark

Measure AI honestly by scoring new models against real human experts rewriting your actual broken work.

Difficulty
Advanced
Time to result
~weeks to results
Steps
3
Confidence
85%

A method for cutting through hyped benchmarks by building your own, grounded in real human expert work. Shipper took a real app that failed in production, had two senior engineers independently rewrite it, and used those rewrites as the ground-truth answer key. He then scores each new model by giving it the same broken code and a prompt, comparing its rewrite against the human baseline. The deeper insight: benchmarks only rise on problems humans have already framed and scored — the framing itself is invisible, unmeasured human work.

Origin

Dan Shipper's 'senior engineer benchmark,' built after his vibe-coded app Proof kept crashing in production and two senior engineers rewrote it independently.

Core principles

  • 01Benchmarks rise only on problems humans have already articulated and scored; the act of framing the problem is unmeasured human work
  • 02AI-vs-human comparisons are really one human-using-AI vs another human-using-AI — AI doesn't use itself
  • 03A good benchmark uses real expert output as ground truth, not synthetic tasks
  • 04Prompts must reveal capability without giving away the answer

How to run it

  1. 1

    Capture a real failure and expert fixes

    Take an actual piece of your own work that broke, and have two independent human experts fix or rewrite it. Their solutions become your ground-truth answer key.

    Pro tip Let the human experts use AI however they like — you're benchmarking human-plus-AI against human-plus-AI, which is the only realistic comparison.

    Watch out Don't hand-write the human baseline without AI; a human who refuses AI isn't a fair or relevant comparison point.

  2. 2

    Craft a prompt that reveals capability without leaking the answer

    Write a prompt that lets the model show what it can really do — e.g. 'this is vibe-coded slop, rewrite it from first principles' — but iterate until the prompt no longer hands the model the solution.

    Pro tip Test the model's agency: good prompts reveal whether it will rip out and rewrite bad code versus papering over the edges with a small patch.

    Watch out A naive prompt like 'here are the issues, go fix them' measures compliance, not judgment — models take it literally and patch instead of confronting the real problem.

  3. 3

    Score each new model against the human baseline

    Run every new model on the same task and score its output out of 100 against the senior-engineer rewrites, tracking the gap and the trajectory over model generations.

    Pro tip Re-run on every release — turning the rock over again is how you catch a capability jump (his score moved from ~30 to 62 across a model generation).

    Watch out A rising score doesn't mean full human replacement; be ready to redesign the benchmark to a higher frame once a model saturates it, because there's always a higher frame of judgment humans add.

In the wild

The Proof rewrite benchmark

Shipper's vibe-coded app Proof crashed every few minutes after launch and he couldn't fix it, so he had two senior engineers rewrite it independently. He then fed the broken code plus a first-principles prompt to each new model. Models scored ~30/100 until one generation jumped to 62, versus human seniors scoring high-80s/low-90s.

A concrete, honest measure of the model-vs-human gap and its rate of closing, grounded in real work rather than a synthetic leaderboard.

Common mistakes

Trusting a benchmark's name over what it measures

A benchmark called 'senior engineer' primes you to read saturation as full replacement, but it only captures pre-framed, scorable problems — the human work of deciding what to build and whether it's good goes unmeasured.

Using a prompt that measures compliance, not judgment

Telling a model 'here are the bugs, fix them' makes it dutifully patch symptoms; you never see whether it has the judgment to recognize the codebase needs a first-principles rewrite.

Is it for you?

Best for

Operators and technical leaders who want an honest, situation-specific read on model capability instead of relying on public leaderboards

Not ideal for

People who need a standardized, comparable benchmark across organizations rather than a bespoke one tied to their own work

From the transcript

I made this senior It's called a the senior engineer benchmark, and it's like, "How good is AI versus a human engineer?"

40:30

I got actually two different senior engineers to fix it independently. So, I have two different rewrites of the code base

41:30

benchmarks rise on problems that we've framed that we can articulate, that we can score

45:00

we're actually really always talking about one human using AI versus another human using AI cuz AI doesn't use itself

48:00

From the episode

The AI paradox: More automation, more humans, more work

Dan Shipper