lab-forge
Executable teaching-artifact builder for STEM professors — pipeline Stage 2 extension. 5-agent team that builds lab and programming assignments as working packages: per-student synthetic datasets with planted, recoverable ground truth, starter-code scaffolds that run as shipped, autograders with visible/hidden test splits, and reference solutions produced by solving and verified by executing. Triggers on: lab assignment, programming assignment, problem set code, dataset for students, synthetic data, starter code, autograder, solution notebook, Jupyter, 实验作业, 编程作业, 数据集, 实验数据, 起始代码, 自动评分, 自动批改, 参考答案.
YujxZJCN · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add YujxZJCN/teaching-skills --skill lab-forgeAbout this skill
name: lab-forge description: "Executable teaching-artifact builder for STEM professors — pipeline Stage 2 extension. 5-agent team that builds lab and programming assignments as working packages: per-student synthetic datasets with planted, recoverable ground truth, starter-code scaffolds that run as shipped, autograders with visible/hidden test splits, and reference solutions produced by solving and verified by executing. Triggers on: lab assignment, programming assignment, problem set code, dataset for students, synthetic data, starter code, autograder, solution notebook, Jupyter, 实验作业, 编程作业, 数据集, 实验数据, 起始代码, 自动评分, 自动批改, 参考答案." metadata: version: "1.0.0" last_updated: "2026-06-10" status: active pipeline_stage: 2 related_skills: - assessment-architect - submission-auditor - lesson-builder - teaching-pipeline
Lab Forge — Executable Teaching-Artifact Team
Builds the artifacts a STEM assignment is actually made of: the dataset students load,
the scaffold they fill in, the tests that grade them, the solution that proves the whole
thing is doable. Where assessment-architect writes the brief and the rubric, this skill
builds the machinery — and machinery is only real when it runs. Everything this skill
ships was executed in session; verification evidence travels with the package.
Prime rule: executed, not assumed. A starter repo that doesn't compile, a dataset whose planted effect the intended analysis can't recover, an autograder the unmodified scaffold can pass — each is a defect, found by running, not by review. Verification is part of the build, never an optional final step.
Quick Start
Build the Week 6 regression lab for STAT 210 — it's A3 in the passport
给我的机器学习课做一个编程作业:起始代码 + 自动评分
Generate a per-student dataset for the ANOVA lab — 45 students, same difficulty each
Write the autograder for this assignment spec; here's the starter repo
Make 4 variants of last year's pathfinding lab so sections can't share answers
我需要这次实验的参考答案和评分说明
Modes
| Mode | Trigger intent | Output |
|---|---|---|
lab | "Build the lab / programming assignment for…" | Complete package: handout + starter + data + verified solution + grader + verification record |
dataset | "Generate data for…" / per-student data requests | Synthetic dataset(s) with planted, recoverable properties; per-student/per-section variants; professor-only ground-truth doc |
starter-code | "Starter code / scaffold / skeleton for…" | Scaffold repo: structure, stubs with full contracts, fixed vs student zones fenced, runs as shipped |
autograder | "Autograder / tests to grade this" | Test-suite grader: visible tests for students + hidden tests for grading, partial-credit map, submission-auditor-consumable output |
solution | "Reference solution / solution notebook for…" | Solution produced by solving from student-facing materials, verified by execution, + grading notes |
variant | "Make N versions of this lab/dataset" | N difficulty-equivalent variants with the equivalence argument stated and spot-checked |
Mode dispatch rule: a request naming a passport assessment loads that entry; a
standalone request intakes context and offers passport write-back at exit (Passport Iron
Rule 5). autograder and solution without existing student-facing materials redirect
to lab — there is nothing honest to grade or solve against yet. Detect intent in any
language.
Does NOT trigger
| Scenario | Use instead |
|---|---|
| Writing the assignment brief, rubric, or grading criteria themselves | assessment-architect project-brief / rubric |
| Checking student submissions against a standard | submission-auditor |
| In-class ungraded exercises and activities | lesson-builder |
| Full design → materials → assessment run | teaching-pipeline |
Agent Team (5)
| Agent | Role |
|---|---|
lab_designer_agent | Designs the lab arc against outcomes: staging, deliverables, submission contract, honest time estimates; safety notes flagged for physical labs |
dataset_smith_agent | Generates synthetic data with planted properties verified recoverable; seeded per-student variation; professor-only ground-truth documentation |
starter_code_agent | Builds scaffold repos that run as shipped: stubs with full contracts, fenced student-work zones, pinned dependencies, tested cold-start README |
autograder_agent | Builds visible + hidden test suites with partial-credit mapping; validates against the verified solution AND the unmodified starter; output feeds submission-auditor |
solution_verifier_agent | Solves the lab cold from student-facing materials only, executing everything; produces solution + grading notes; reports defects rather than patching around them |
Workflow (lab mode)
Phase 0 INTAKE — load the passport assessment entry (outcomes_assessed, weight,
week, ai_tier) or intake standalone: outcomes, topics taught,
student environment (language, tools, compute), class size.
Missing context = ask, don't guess (Passport Iron Rule 2).
Phase 1 DESIGN — lab_designer drafts the lab arc: what students do, in what
order, what they submit (the submission contract), time
estimate with evidence
🧑 checkpoint: arc confirmed — staging, deliverables, and per-student
variation plan decided here, before anything is built
Phase 2 BUILD — parallel: dataset_smith (data + ground_truth.md) and
starter_code (scaffold + env). Both execute their artifacts
before handing off.
Phase 3 SOLVE — solution_verifier solves the lab from the STUDENT artifacts
only — handout, starter, data — executing every step. A lab
that can't be completed from its own materials is repo
Read the complete source ↗