01// Maieutic
02// programming education for the LLM era.
03// 3rd place · Anthropic Claude Code x Cerebral Valley Hackathon · Built with Opus 4.7
An LLM that flips the role it usually plays in learning to program. Instead of handing students working code, it guides them to think first about what the program should do — the specification — before they write a single line.
Writing code from scratch matters less than it used to. Specifying behavior precisely, reading code critically, and noticing the gap between intent and output matter more. Those are the skills Maieutic trains.
Along the way it surfaces a learning signal that's normally invisible to an instructor: not just whether the code passes the tests, but whether the student can explain why their code does what it does, and where they drifted from their own spec.
Before writing any code, the student describes what the program should do — clearly enough that someone else could implement it. Opus reads the description and asks the obvious questions the student left unanswered. The editor stays locked until the spec answers them.
Students write in Monaco with autocomplete off. Opus answers reference questions directly ("syntax for a for-loop?") but returns counter-questions for reasoning ones ("why doesn't my count look right?"). The thinking stays with the student.
On submission, Opus compares the spec to the code. Wherever the two don't line up, it points the difference out as a neutral question — "In your spec you said X. In the code I see Y. What happened?"
One row per active student, each with a one-sentence summary of where they actually are in their thinking — not "phase 3, 6 minutes idle" but "the student wrote 'n >= 0' and 'negative inputs are handled' in the same spec; they're confused about committing to behavior, not about Fibonacci."
For any session: on the left, everything the student wrote. On the right, instructor-only, what the system understood about that student — what it predicted they would say, how their actual answer compared, where the gap is.
After a class works through a problem, a short narrative of how the cohort did — not "students struggled" but "six of eight missed case-sensitivity on their first spec; consider introducing it as an explicit dimension earlier in the unit."
Every published exercise becomes a card with aggregate results — completion rates, which spec dimensions get missed most, how divergences distribute. And the authoring side: a plain-text problem prompt becomes spec scaffolding the teacher can review and publish in seconds. A unit's worth of exercises is an afternoon, not a week.
Anthropic Claude Code x Cerebral Valley Hackathon
"Built with Opus 4.7" track.
For pilots, collaboration, or questions about Maieutic.
hola@maieutic.dev