Skip to content

CLAUDE.md

Latest commit

 

History

History
87 lines (63 loc) · 3.58 KB

CLAUDE.md

File metadata and controls

87 lines (63 loc) · 3.58 KB

MTAT — Modular Training Architecture Template

A Python CLI system for AI-assisted adaptation of training course modules. One person maintains base content; Claude adapts it for unlimited audiences and locales.

Project Overview

  • generate-variant.py — Core CLI: reads a module (base.md + metadata.yaml), calls Claude API, writes an audience-adapted variant to variants/
  • upload-to-moodle.py — Uploads generated variants to a local Moodle LMS via Docker for preview
  • prompts/adapt.md — Version-controlled system prompt governing Claude's adaptation behavior
  • RUBRIC.md — Quality scoring rubric (1–5 per dimension) for evaluating base modules and variants
  • example-course/ — Working example: Prompt Engineering Fundamentals (4 modules: concept, demo, exercise, assessment)
  • variants/ — Generated output (gitignored except manifest.yaml)

Environment Setup

python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt
export ANTHROPIC_API_KEY=your_key_here   # or add to .env

Key Commands

# Generate a variant
python3 generate-variant.py --module example-course/01-concept --audience developer
python3 generate-variant.py --module example-course/02-demo --audience executive --locale es-MX

# Start local Moodle (Docker)
docker compose up -d

# Upload variants to Moodle (first time)
python upload-to-moodle.py --setup-moodle
# Subsequent runs
python upload-to-moodle.py

Audience Presets

Preset Who they are
developer Engineers building with LLMs/APIs — technical, mechanism-first
executive Senior leaders — ROI framing, business outcomes
champion Internal trainers — facilitation notes, timing, debrief prompts
technical-writer Doc professionals — structured authoring analogies

Custom audiences are supported: --audience "healthcare compliance officers with no prior AI experience"

Module Structure

Each module is a directory with exactly two files:

01-concept/
├── base.md          # Content — plain Markdown, audience-agnostic
└── metadata.yaml    # Structured data: ID, type, objectives, version, tags

Module types: concept | demo | exercise | assessment

Adding a New Course

  1. Create a subdirectory (e.g., my-course/)
  2. Add numbered module subdirectories with base.md + metadata.yaml
  3. Score each module against RUBRIC.md — must score ≥3 on all dimensions before publishing
  4. Generate variants: python3 generate-variant.py --module my-course/01-concept --audience developer

Variant Manifest

Every generation appends to variants/manifest.yaml — the audit trail tracking module ID, audience, locale, output file, timestamp, model, and token counts (for cost visibility).

Architecture Notes

  • base.md has no front matter — the script prepends YAML front matter to each generated variant
  • prompts/adapt.md is version-controlled — changes affect all future variants; commit with a clear rationale message
  • Generated variant files are gitignored; only manifest.yaml is committed
  • Add audience presets by extending AUDIENCE_PROFILES dict in generate-variant.py
  • Module IDs (e.g., pe-fundamentals-01) are stable across versions and used as manifest keys

Dependencies

  • anthropic>=0.40.0 — Claude API client
  • pyyaml>=6.0 — metadata parsing and manifest writes
  • python-dotenv>=1.0 — .env file support for API key
  • requests>=2.31.0 — Moodle REST API calls
  • markdown>=3.5.0 — Markdown-to-HTML conversion for Moodle upload