AI coding assistants should make you think before they give you the answer

The dominant use of AI coding assistants is to skip the hard part: you describe what you want, the model writes it, you paste it in. This toolkit inverts that contract deliberately. The AI is not allowed to hand you a solution until you have committed to a prediction, articulated what you expected, and attempted an explanation. The friction is the point.

Twelve slash commands structure the interaction. /hint is the workhorse — you state what you think should happen, then the AI releases the smallest possible nudge, escalating only if you keep asking. /debug refuses to name the bug until you have formed a hypothesis about what went wrong. /autopsy runs after a fix, not before, forcing a post-mortem that classifies the error and asks what belief you held that turned out to be false. /explain is a teach-back test: you explain a concept as if to a beginner, and the AI probes for vague terms and contradictions rather than accepting a plausible-sounding summary.

Four reusable "skills" — debugging, examination, code-review, retrieval — shape how the AI responds across prompts without being invoked directly. They encode specific epistemic disciplines: the debugging skill requires a hypothesis before any cause is suggested; the examination skill corrects only the smallest misconception first rather than delivering a lecture. These are not personality settings; they are behavioral constraints on the model's response pattern.

The optional learning logs folder — four markdown files for mistakes, concepts, open questions, and review metadata — feeds the /retrieve command, which reads past entries and generates spaced-retrieval questions mixing prediction, debugging, and application. That loop closes the system: what you got wrong last week becomes a question this week.

The escape hatch is honest. Saying "ship this" switches the AI to normal engineering mode. The learning constraints only apply when you invoke a learning prompt, which means the toolkit can coexist with real work without forcing pedagogy on a deadline.

The whole thing is a folder of markdown files and prompt definitions — no dependencies, no scripts. It requires GitHub Copilot Chat with inline completions disabled, and the skills layer needs a Copilot plan that supports custom skills. That last constraint is worth noting: the prompts work on any plan, but the cross-prompt behavioral consistency depends on the skills feature being available.

What this gets right is the mechanism. Cognitive science on learning is fairly clear that generation and retrieval practice beat passive reading, and that feedback timed to a committed answer is more durable than feedback delivered before the learner has formed a belief. This toolkit operationalizes those ideas inside the tool most developers already have open. Whether the prompts hold up under real use depends on how well Copilot follows the constraints — a model that drifts toward helpfulness will leak answers — but the design intent is sound.