dsh-project-based-learning

Most AI tutoring sessions share a common flaw: the learner says "I get it" and the coach moves on; one correct diagnostic question earns a "mastery" label. Factual knowledge and procedural skill are fed together, AI-written code counts toward competence, and stage acceptance rests on "looks about right". Subject content is hardcoded into prompts, so switching topics means starting over.

dsh-project-based-learning splits the problem in two: pedagogy lives in the engine, subject knowledge lives in swappable domain packs. The engine defines the protocol — phased roadmaps, task cards, evidence archives, three-tier acceptance verdicts — while the domain pack supplies content: project archetypes, diagnostic banks, verification recipes, and pitfall catalogs. Switching subjects never touches the engine; the first bundled domain pack covers Unity / C#. Crucially, rules like "reject competence claims without supporting evidence," "block stage passage when unresolved obstacles exist," and "no subject terms in engine files" are not wishful prompts — they are enforced by a zero-dependency validator that runs mechanically after every checkable action. State lives solely in .coach/state.json; PROGRESS.md is a rendered view, not editable by hand.

If you are working through a concrete technical track and refuse to let an AI write the code and pass it off as your skill, this plugin is built for that mode of learning. It drives you to build, forces you to explain, demands you transfer to a new context, and only then records the result in your archive. The coach runs the loop; the learner does the real work.