Generate Lectures

Iterative Lecture Generator

Generate complete lecture packages from a course outline by running the VectorShift Individual Lecture HTML pipeline (with SVG diagrams) for each lecture sequentially.

Slash Command Usage

/generate-lectures <outline_path> [options]

Examples:

• /generate-lectures outline.md - Generate all lectures
• /generate-lectures outline.md --dry-run - Preview without API calls
• /generate-lectures outline.md --start 2 --end 4 - Generate lectures 2-4 only
• /generate-lectures "Abid Husain/course_outline.md" --materials "Abid Husain/" - Custom paths

Execution Instructions

When this skill is invoked, execute the following steps:

1. Parse arguments from the skill invocation
2. Run dry-run first (unless user explicitly skipped it) to show the execution plan
3. Confirm with user before making API calls
4. Execute the generator:

python3 "/Users/anantvinjamoori/Vectorshift Pipelines/cli/iterative_lecture_runner.py" \
    "{outline_path}" \
    {--dry-run if specified} \
    {--start N if specified} \
    {--end N if specified} \
    {--materials PATH if specified}

When to Use This Skill

Use this skill when the user:

• Has a course outline (markdown) and wants to generate all lectures
• Says "generate lectures", "create the course", "run the lecture pipeline"
• Asks to "batch process" or "iterate through" lectures
• Provides a course outline and materials folder
• Invokes /generate-lectures

Pipeline Information

Required Inputs

1. Course outline file - Markdown file with lecture structure
2. Materials folder (optional) - Folder with PDF materials (default: Abid Husain/)

Workflow

Step 1: Gather Information

Ask the user for:

• Path to course outline markdown file
• Materials folder path (or confirm default: Abid Husain/)
• Lecture range to process (start/end numbers, optional)
• Whether to do a dry run first

Step 2: Run the Helper Script

Execute the Python helper:

python3 "/Users/anantvinjamoori/Vectorshift Pipelines/cli/iterative_lecture_runner.py" \
    "{course_outline_path}" \
    --materials "{materials_folder}" \
    --start {start_lecture} \
    --end {end_lecture}

Options:

• --dry-run - Preview without making API calls
• --start N - Start from lecture N
• --end N - Stop after lecture N
• --materials PATH - Custom materials folder
• --output PATH - Custom output folder
• --no-llm-parser - Disable LLM parser and use regex instead (LLM parser is default)
• --claude-api-key KEY - Claude API key for LLM parser (uses env var if not provided)

Step 3: Monitor Progress

The script will:

1. Parse the course outline to find lectures (regex or LLM parser)
2. Extract suggested materials for each lecture
3. Fuzzy-match material names to PDF files
4. Submit job to VectorShift with background: true (async)
5. Poll for results every 5 seconds (max 30 minutes per lecture)
6. Save outputs to outputs/{course_name}/

Note: The async API pattern eliminates HTTP timeout issues for long-running pipelines.

Step 4: Report Results

After completion, show the user:

• Number of successful/failed lectures
• Location of output files
• Any error messages

Output Files

For each lecture, the following files are saved to outputs/{course_name}/:

Course Outline Format

The skill expects course outlines with this structure:

# Course Title

## Lecture 1 - Introduction to Topic

### Where to integrate user-provided materials
- Use "BPC-157 and the Cardiovascular System" dossier
- Reference "SS-31 in Cardiovascular Medicine" protocols

### Suggested preparatory materials
- Reading: "Apolipoprotein B: Bridging the Gap"

---

## Lecture 2 - Deep Dive
...

Material Matching

The skill fuzzy-matches material references to PDF files:

Available Materials (Abid Husain folder)

49 PDF files including:

• BPC-157 research papers
• SS-31/Elamipretide studies
• GLP-1 and cardiovascular effects
• Testosterone therapy research
• Peptide therapy protocols
• Cardio-Zoomer documentation
• Cardiovascular risk assessment guides

Error Handling

Output Retrieval Fallback

If output capture fails locally (timeout, network issues, interrupted process), the task_id fallback triggers automatically—no manual intervention required.

How It Works

VectorShift confirmed: The task_id returned when submitting a job can be used directly to query results. This eliminates the need to manually fetch span IDs from the UI.

task_id from job submission == span_id for status queries

Automatic Recovery

The runner scripts now automatically:

1. Store the task_id from each job submission
2. Use the task_id to retry fetching results on timeout
3. Log all task IDs for manual recovery if needed

Manual Recovery (if needed)

If you have a task_id from a previous run:

# Use task_id directly (same as span_id)
python3 "/Users/anantvinjamoori/Vectorshift Pipelines/cli/fetch_by_span_id.py" \
    69601d086fdec16163dc80fe \
    <TASK_ID> \
    --output-dir ./output

Or use Python:

from vs_pipelines.config import fetch_pipeline_result_by_span_id

result = fetch_pipeline_result_by_span_id(
    pipeline_id="69601d086fdec16163dc80fe",
    span_id=task_id  # task_id works directly!
)

if result["status"] == "completed":
    slides = result["result"].get("lecture_json", "")
    transcript = result["result"].get("transcript", "")

See vectorshift-pipeline-deployment.md for full documentation.

Example Usage

Via slash command:

/generate-lectures outline.md
/generate-lectures outline.md --dry-run
/generate-lectures outline.md --start 2 --end 4
/generate-lectures outline.md --no-llm-parser  # Use regex instead

Via natural language:

User: "I have a course outline at outline.md. Generate all the lectures."

Claude: First runs dry-run to show plan, then confirms with user before executing. LLM parser (Claude Haiku 4.5) is used by default.

User: "Generate just lectures 2 through 4"

Claude: Runs with --start 2 --end 4 flags.

User: "Use the regex parser instead"

Claude: Runs with --no-llm-parser flag to disable LLM parsing.

Execution Notes

• Each lecture takes 5-10 minutes to process
• Total time for 5 lectures: ~30-50 minutes
• Progress is displayed in real-time
• Failed lectures don't stop the process
• Outputs can be resumed with --start N