Teaching Skills for Claude Code

English | 简体中文

A comprehensive suite of Claude Code skills for university professors, covering the full teaching lifecycle: design → build → assess → deliver → reflect → improve.

15 skills · 84 modes · 71-agent ensemble · 2 quality gates · 1 Course Passport

AI is your teaching assistant, not your replacement. This suite won't teach for you. It handles the structure and the grunt work — drafting outcomes, building blueprints before exams, auditing alignment, formatting syllabi, coding evaluation comments — so you can focus on what actually requires you: knowing your discipline, knowing your students, and deciding what matters. Every consequential decision passes through a checkpoint where you decide.

Architecturally inspired by Academic Research Skills (Cheng-I Wu) — the research-side sibling of this suite. If you also write papers, the two compose: teaching-reflector's SoTL mode hands off to ARS's deep-research and academic-paper skills.

Install

Plugin (recommended):

/plugin marketplace add YujxZJCN/teaching-skills
/plugin install teaching-skills

Manual: clone this repo and symlink the skill directories into ~/.claude/skills/ (or the project's .claude/skills/).

Verify: run /ts-status in a directory with a course_passport.yaml, or just say "Design a new course on X for 60 second-year students."

The skills

Extension skills — production, operations, and compliance layers; each works standalone and the pipeline dispatches them on demand:

The pipeline

Stage 0 CONTEXT  → Course Passport initialized                 🧑
Stage 1 DESIGN   → course-designer                             🧑
Gate 1.5 ALIGNMENT — constructive-alignment audit              ✓ machine + ack
Stage 2 BUILD    → lesson-builder (just-in-time by default)    🧑
Stage 3 ASSESS   → assessment-architect + integrity audit      🧑
Gate 3.5 QUALITY — AI-integrity, transparency, UDL, workload   ✓ machine + ack
Stage 4 DELIVER  → weekly loop: materials · mentoring · submission audits · midcourse feedback
Stage 5 REFLECT  → teaching-reflector eval-analysis            🧑
Stage 6 IMPROVE  → iteration record → next term re-enters Stage 1

🧑 = professor checkpoint (you decide) · ✓ = deterministic gate, then your acknowledgment

Two ideas hold this together:

• The Course Passport (shared/course_passport_schema.md) — a YAML file that is the single source of truth for the course. Outcomes, assessment plan, schedule, policies, gate findings, artifact ledger, iteration history. State lives in the passport, not the conversation: any fresh session resumes from it. Every skill also works standalone without one.
• Constructive alignment as a machine check — the Alignment Gate verifies the outcome–teaching–assessment triangle is closed before anything is built on it; the Quality Gate verifies the built artifacts are transparent (TILT), accessible (UDL), integrity-coherent (AI-era), and humanly workloaded — before students see them.

Showcase: what the output looks like

Two complete, synthetic-but-rule-obeying artifact sets — so you can judge output quality before installing:

• STEM lecture course → — CS 304 Introduction to Machine Learning, 90 students: filled Course Passport, syllabus, alignment gate report with a dismissed-warning trail, lesson plan + activity sheet, exam blueprint with sample items and a worked key, the AI-integrity audit that redesigned the course project, and a rendered HTML course dashboard.
• Humanities seminar → — HIST 247 The Atlantic World, 18 students: a discussion/writing course with specs-flavored grading, plus inspectable person-affecting outputs — a recommendation letter with its claim-to-intake trace appendix, the intake ledger behind it, and a feedback report — demonstrating the anti-fabrication templates (every claim traces to evidence; no student data in the passport).

Quick start

# Full lifecycle
"I'm teaching a new course on environmental economics next fall — run the full pipeline"

# Blank page
"Guide me — I know my field but I've never designed a course"          → socratic

# Single artifacts
/ts-outcomes  /ts-syllabus  /ts-lesson  /ts-exam  /ts-rubric  /ts-letter  /ts-evals

# Mid-entry
"I already have a syllabus, build me week 3's materials"
"Semester's over, here are my student evaluations"                      → eval-analysis

# Status
/ts-status    "What's next for my course?"

See QUICKSTART.md for a worked first session and MODE_REGISTRY.md for all 84 modes.

Design principles

1. The professor is the pilot. Checkpoints at every stage; "key decisions made for you" are always surfaced, never buried; "just proceed" is respected (shared/checkpoint_protocol.md).
2. Evidence-based defaults, professor's final call. Recommendations cite shared/pedagogy_foundations.md (backward design, constructive alignment, active learning, retrieval practice, TILT, UDL, cognitive load, feedback research). When you overrule a principle, the decision is logged and never re-litigated.
3. No invented context. Empty learner profile → the skill asks. Institutional policies → [NEEDS PROFESSOR INPUT] markers, never plausible filler. Uncertain domain claims in generated materials → [VERIFY] markers.
4. Person-affecting outputs are evidence-bound. Feedback, letters, and intervention plans use only material you provided — no invented anecdotes, ever — and always end with a verify-before-send reminder. Student data never enters the Course Passport.
5. Integrity by design, not detection. The AI-era integrity model (shared/ai_era_integrity.md) uses per-assignment policy tiers and structural resilience patterns. Nothing in this suite relies on AI-detection tools.
6. The contracts are machine-checked. The Course Passport has a JSON Schema (shared/course_passport.schema.json) and validators: scripts/check_passport.py (cross-reference invariants — id mirrors, weight sums) and scripts/check_alignment_gate.py (Gate 1.5 executed deterministically, not interpreted by a model). CI runs them plus a mutation-test suite on every push.
7. Honest evidence reading. Student evaluations are treated as biased, small-sample evidence of student experience — analyzed thematically, triangulated, and caveated — not as a teaching-quality score.

Repository layout

course-designer/        SKILL.md + agents/ + references/ + templates/
lesson-builder/         〃
assessment-architect/   〃
student-mentor/         〃
submission-auditor/     〃
teaching-reflector/     〃
teaching-pipeline/      SKILL.md + agents/ + references/
deck-studio/            extension skills, same layout
lab-forge/              〃
media-scripter/         〃
course-publisher/       〃
ta-coordinator/         〃
accreditation-mapper/   〃
bilingual-courseware/   〃
cohort-analyst/         〃
shared/                 Course Passport schema (prose + JSON Schema) · pedagogy
                        foundations · gate protocols · AI-era integrity · checkpoints
scripts/                check_passport.py · check_alignment_gate.py · build_dashboard.py · registry lint
tests/                  validator suite (golden fixture + mutation tests)
commands/               /ts-* slash commands
docs/                   ARCHITECTURE.md
.claude-plugin/         plugin + marketplace manifests
skills/                 symlinks for plugin auto-discovery

Languages

English and Simplified Chinese trigger keywords ship by default; intent-based modes (socratic design, mid-entry routing) work in any language. Add your language's keywords to a skill's description frontmatter to improve trigger matching.

Portability

This repo is the Claude Code distribution (native skills, /plugin install, /ts-* slash commands). The suite is three layers with different reach:

• Skills (the SKILL.md content) follow Anthropic's Agent Skills format and also work on other Claude surfaces (claude.ai, Claude Desktop, the IDE extensions) via the Skills feature.
• The Python tooling (scripts/, the JSON Schema, the dashboard) is fully tool-agnostic — it runs in any terminal with python3, no AI required.
• The plugin packaging (marketplace install, slash commands) is Claude-Code-specific.

For OpenAI Codex CLI, a sibling distribution is generated from this repo by scripts/build_codex.py and published at YujxZJCN/teaching-skills-codex — the whole suite as a single Codex skill with ts-* aliases. It is a build product (regenerated, never hand-edited), so it never drifts from this source.

License

MIT. The architecture (skills/agents/gates/passport pattern) is inspired by Academic Research Skills (CC BY-NC 4.0); no content from that project is reproduced here.