dsh-human-coding

After letting AI generate code for long stretches, you start to feel like a reviewer rather than a coder. dsh-human-coding puts the AI's hands back in its pocket and keeps only the brain: you state a requirement, AI scaffolds the structure, writes the tests, and pops a challenge card. Then the roles flip — you fill in every implementation while AI stays in the back seat, answering questions and handing out hints. Each hint gets more specific than the last, escalating from direction to approach to near-answer, but the budget is finite and AI will never write the code for you. It is not about whether AI can code for you; it is about whether you still can.

Once you submit, AI runs the tests, reviews your implementation line by line, and assigns a completion score from 0 to 1. You can keep iterating without spending hint budget, or give up entirely. When the score passes, you decide the final code: keep your implementation as-is with AI's commentary, let AI make small tweaks on top of yours with per-hunk explanations, or let AI do a full rewrite and walk through the differences. There is a guardrail: challenges only trigger when the request involves real code that has been scaffolded into project files — pure Q&A or single-line edits never fire. If the scaffold itself is broken, AI voids the question on the spot, rolls back all counters, and leaves zero traces.

Every challenge is logged into persistent statistics: acceptance rate, pass rate, average submissions, hint consumption, finish distribution, average completion, and a per-question performance score (completion × difficulty × 100, top-hundred weighted by a 0.95 geometric decay and normalised to 100). The settings panel lets you page through individual records, sort by contribution, reset with one click, and export or import JSON with either overwrite or merge-on-duplicate semantics for easy machine migration. Question state survives restarts via plugin-level persistence and never writes extra files into your project. Ideal for developers who want to keep their coding touch without fully abandoning AI assistance, or anyone who simply enjoys the ancient pleasure of being quizzed and graded.