Try it
The problem
Reading tips about prompt engineering doesn’t build the skill — people nod along, feel like they get it, then write the same weak prompt next time. What actually works is being forced to produce something, have it critiqued, and try again. But standing up a real training platform usually means a backend, accounts, and a grading service before a single lesson ships.
What I built
A belt-ranked prompt-engineering dojo — white through shodan, nine belts — as one self-contained HTML file. Each belt pairs a short concept explanation and a worked “miss/fix” example with three graded exercises: a kata (drill), a sparring round (diagnose and rewrite someone else’s broken prompt), and a test that gates progression to the next belt.
How it works
- No backend, no API key, no build step. Pure HTML, CSS, and vanilla JS — it runs by opening the file.
- Progress lives in the browser. Belts unlock sequentially via
localStorage; nothing is sent anywhere. - Grading works by handoff, not API integration. The tool builds a grading prompt around whatever the learner wrote, copies it to the clipboard, and opens an AI chat interface. The learner reports the verdict back into the tool, which unlocks the next belt on a pass. No API costs, and it works with whatever AI access the learner already has.
What’s next
I’ve since scoped a v2 that generalizes the same belt/kata/sparring format beyond prompt engineering to any topic with a machine-gradable mastery signal, and replaces the self-report grading step above with automated grading. That’s a requirements document at this stage, not shipped code — stress-testing the idea against several very different topics (chosen specifically to find where automated grading breaks down, like judging tone in audio) surfaced real technical and fairness limits worth scoping around deliberately rather than glossing over.
Stack
Vanilla JavaScript, CSS custom properties, LocalStorage. No backend, no build step, no dependencies.