ProMentor

What it is in one line

An agent skill that turns any real open-source project into a step-by-step hands-on course: install it into Claude Code / Codex / DeepSeek Harness and the AI stops writing code for you and starts making you write it yourself, then grades you.

Who built it

A personal open-source project by Lyn-77, repository created 2026-07-31, 41 commits. README, command reference, architecture notes, and a release-packaging flow are all polished to near-product level. No license file.

Read: a solo author who took engineering completeness to product level — command system, web dashboard, plugin registration. But the author's background, target users, and any evidence of real use are all undisclosed.

What it actually does

• Generates a course → /promentor init scans the project architecture and produces a curriculum outline; once confirmed, it generates lectures, labs, and behavioral tests chapter by chapter
• Staged learning → /promentor learn delivers the lecture, walks you through annotated source, and guides you to hand-write the core logic
• Behavioral-test grading → /promentor test runs behavioral tests and tells you what passed, what failed, and why
• Layered hints → /promentor hint reads your code and test results and gives layered hints for the exact error — direction first, then approach, never the answer
• Submit and review → /promentor submit runs the full test suite and locks your score; /promentor review diffs your implementation against the original source and explains the design decisions
• Progress and dashboard → /promentor progress for totals; a web dashboard (Next.js) shows completion, chapter states, and scores

The learning model is a chain: learn the concept, read the source, implement the lab, run the tests, submit and review, master the system design.

What old behavior it replaces

Three old paths for learning to code from a real codebase:

• Reading the source directly → random jumping with no dependency order; ProMentor generates a dependency-ordered path
• Watching video courses → watching someone else write; ProMentor requires you to hand-write the core logic
• Blogs and docs → fragmented knowledge points; ProMentor requires a full understanding of one system's design philosophy

For the agent itself, it replaces passive Q&A ("explain this repo to me") with an active, tested, graded learning loop. The critical transfer is: grading does not come from the AI eyeballing answers, it comes from behavioral tests.

Business model

Not disclosed. Free and open source, no license declaration, no paid tier, no cloud service.

Read: skill packs are a zero-barrier category — anyone can write one, and they are naturally hard to charge for. The real competition is not distribution; it is whether the behavioral-test grading design holds up on projects at real scale.

Hard numbers

• 48 stars, 1 fork, 0 open issues (fetched 2026-08-14). Created 2026-07-31, 41 commits
• Release zip ~3.7 MB (skill + prebuilt dashboard + plugin)
• No license file; no usage numbers, no user evidence
• Not seen on HN; stars from organic GitHub traffic

Four-way read

The call

The pattern is interesting; the category is cheap. "Behavioral-test grading + layered hints without giving answers" is one of the few genuinely pedagogical designs in this batch — grading comes from executable tests, not a model's subjective score, and hints escalate from direction to approach. Both are directly copyable by anyone building an "AI tutor" product.

The ceiling is scale. Generating discriminating behavioral tests works on a small project. Can it still produce tests with real discriminative power on a multi-ten-thousand-file repository? That is its true scale ceiling, and the README does not answer it.

Risk list: solo author, no license, no user evidence. Skill packs are easy to copy and hard to moat.

What to watch next

① Star growth, and whether anyone publicly reports completing a full course on a real project ② How large a codebase behavioral-test grading still holds up on (the scale ceiling) ③ Whether a license appears and whether contributors show up (survival signals)

What you can take from it

Product logic: three pieces for teaching AI products — generate a staged course from a real artifact (project source); grade with behavioral tests instead of letting the AI score against a reference answer; give layered hints (direction, then approach, never the answer). Directly usable for "AI tutor" or "generate a course from your data" products.

Pricing structure: none. Free and open source.

Verdict

Unproven. The teaching design has incremental value, the category has a shallow moat, and validation is thin. Come back in three months against the three checks above.