The problem
Practice platforms grade the final query or hand over a hint when you’re stuck. Neither handles the real gap: it’s hard to picture what you’re starting with and where the query needs to end up, before writing a single line.
The approach
A coaching skill built around one rule: showing the correct query is the failure case, not the fallback.
How it works
- Every session runs the same four-step pre-flight first: output shape, source, filters, aggregation
- If stuck, it climbs a hint ladder from least to most revealing, so the learner writes each clause themselves
- Once there’s a working query, it’s explained visually: table transformations, join diagrams, an annotated breakdown
Design decisions
- Even “just show me the answer” gets redirected, treated as needing to get unstuck, not a shortcut
- The four-step pre-flight wasn’t planned: an early version had me redo a query blind and report what failed, and when I said the real gap was picturing what I had and where I was going, it built the pre-flight to close that
What this demonstrates
The checks that move are what make someone independent afterward: decomposing before writing SQL, working through a stuck point instead of being handed it. The checks that don’t move are already table stakes for a good model. The skill isn’t teaching SQL, it’s teaching the habit of decomposition.
100% vs. 50% pass rate, one exercise, 3 runs per configuration. Full run data in benchmark.md.
Designed instruction, not a chatbot with a personality: a named teaching philosophy and an escalation rule that holds even under pressure to skip it.
