Podcast Producer: End-to-End Episode Orchestration

• "Create an episode for [Drug] and [Disease]"
• "Produce the [Drug]-[Disease] episode"
• "Generate podcast for [Drug] treating [Disease]"
• "Make an episode about [Drug] for [Disease]"

Normalize to format: pair-<drug>-<disease> (lowercase, hyphens for spaces)

Step 2: Verify Working Directory

Ensure the working directory structure exists:

stories/pair-<drug>-<disease>/
├── background/      (created by archivist)
├── shownotes/       (created by showrunner)
└── transcript/      (created by podcast-writer)

Step 3: Run Production Pipeline

Execute agents sequentially using the Task tool. After each agent:

1. ✅ Verify output (check expected files exist)
2. ✅ Auto-cleanup (fix common naming/location mistakes)
3. ⚠️ Retry once if issues remain (with specific feedback)
4. ❌ Fail if still broken after retry

Execute agents in this order:

Agent 1: Archivist (Research)

Task tool parameters:
- subagent_type: "archivist"
- description: "Research background for [drug]-[disease]"
- prompt: "Research and compile a comprehensive, evidence-first background dossier for
           [Drug Name] repurposed for [Disease Name]. Use specialized sources first
           (PubMed, ClinicalTrials.gov, bioRxiv/medRxiv, patents, FDA/EMA docs), then
           general web search only to fill gaps. Generate detailed background files
           with citations covering context, origin, evidence of discovery/trigger,
           regulatory/adoption history, mechanism, and impact."

Output: stories/pair-<drug>-<disease>/background/*.md files

Verification: Check that at least 5 background files were created with substantial content

Agent 2: Showrunner (Structure)

Task tool parameters:
- subagent_type: "showrunner"
- description: "Create episode structure for [drug]-[disease]"
- prompt: "Using the background dossier at stories/pair-<drug>-<disease>/background/,
           choose the narrative structure that best fits the evidence, then create the episode
           structure with cold open, chapter outline, thematic playbook, and grading criteria.
           Output to stories/pair-<drug>-<disease>/shownotes/"

Output: stories/pair-<drug>-<disease>/shownotes/*.md files

Verification: Check that 00-episode-structure.md exists with narrative arc and timing

Agent 3: Podcast-writer (Scripting)

Task tool parameters:
- subagent_type: "podcast-writer"
- description: "Generate transcript for [drug]-[disease]"
- prompt: "Using the episode structure at stories/pair-<drug>-<disease>/shownotes/,
           generate full podcast transcript with dual-host dialogue (Marcus & Elena),
           following the showrunner's chosen structure. Ensure equal speaker distribution
           and varied emotional tags. Format for ElevenLabs podcast API. Output to
           stories/pair-<drug>-<disease>/transcript/"

Output: stories/pair-<drug>-<disease>/transcript/*.md files (one per chapter)

Verification: Check that multiple transcript sections exist with balanced speaker distribution

Agent 4: Editor (Editorial Pass)

Task tool parameters:
- subagent_type: "editor"
- description: "Editorial pass for [drug]-[disease]"
- prompt: "Perform comprehensive editorial pass on transcript at
           stories/pair-<drug>-<disease>/transcript/. Fix narrative flow, transitions,
           duplication, pacing issues. Update transcript files directly and create
           editorial summary."

Output: Updated transcript files + editorial notes

Verification: Check that editorial notes or summary file was created

Agent 5: Speaker (Audio Generation)

Task tool parameters:
- subagent_type: "speaker"
- description: "Generate audio for [drug]-[disease]"
- prompt: "Generate podcast audio from completed transcript at
           stories/pair-<drug>-<disease>/transcript/ using Google Gemini TTS API.
           Output MP3 file to stories/pair-<drug>-<disease>/transcript/"

Output: stories/pair-<drug>-<disease>/transcript/*.mp3 file

Verification: Check that MP3 file exists and has non-zero size

Step 4: Git Commit

After successful completion, commit all work:

git add stories/pair-<drug>-<disease>/
git commit -m "Complete episode production: <Drug> - <Disease>

- Research: Background dossier compiled by archivist
- Structure: Episode outline and show notes created
- Script: Dual-host transcript generated
- Editorial: Final narrative pass completed
- Audio: Full episode MP3 generated

Episode ready for review and publishing.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <[email protected]>"

Important: Do NOT push to remote. Only commit locally.

Step 5: Report Completion

Provide summary to user:

• Episode location
• Files created
• Git commit hash
• Next steps (review audio, gather feedback, publish)

Error Handling

If any agent fails during the pipeline:

1. Create Error File

Write detailed error report to:

stories/pair-<drug>-<disease>/PRODUCTION_ERROR.md

Error file format:

# Production Error: [Drug] - [Disease]

**Date**: [ISO timestamp]
**Failed Agent**: [Agent name that failed]
**Pipeline Stage**: [Stage number and name]

## Error Description

[Detailed error message and stack trace if available]

## Work Completed

- [x] Archivist: Background dossier completed
- [x] Showrunner: Episode structure completed
- [ ] Podcast-writer: FAILED HERE
- [ ] Editor: Not started
- [ ] Speaker: Not started

## Files Created

[List of files successfully created before failure]

## Required Action

TODO: Human intervention required to resolve this error and continue production.

[Specific guidance on what needs to be fixed based on the error]

## Resuming Production

After fixing the error, resume from the failed agent:
[Command to resume, e.g., "Continue podcast production from podcast-writer step"]

2. Commit Partial Work

Even on failure, commit what was completed:

git add stories/pair-<drug>-<disease>/
git commit -m "PARTIAL: Episode production incomplete: <Drug> - <Disease>

Pipeline failed at [Agent Name] stage.
See PRODUCTION_ERROR.md for details.

Work completed:
- [List completed stages]

TODO: Human intervention required to resolve production error.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <[email protected]>"

3. Report Failure to User

Clearly communicate:

• Which agent failed
• What was completed successfully
• Location of error file
• How to search for TODO markers: grep -r "TODO" stories/pair-<drug>-<disease>/

User Control Options

Stop at Specific Step

If user specifies stopping early:

• "Create episode for X and Y, but stop after showrunner"
• "Only run archivist for X and Y"
• "Generate up to transcript for X and Y"

Then run only the specified agents and stop. Still commit the partial work.

Skip Completed Steps

If directory already has completed work:

• Check which agents have already produced output
• Ask user: "I see [X agent] already completed. Resume from [Y agent]?"
• Or if user is explicit: "Continue from editor step"

Re-run Specific Agent

If user wants to re-run a single agent:

• "Re-run the podcast-writer for X and Y"
• "Regenerate audio for X and Y"

Then run only that agent, potentially overwriting previous output.

Quality Checks & Cleanup

After each agent completes, perform verification and automatic cleanup:

Verification Process

1. Check expected files exist (see checklists below)
2. Check file content (not empty, no error messages)
3. Detect file structure violations (wrong names, wrong locations)
4. Automatically fix common mistakes (rename, move files)
5. Report what was cleaned up to user

Agent-Specific Checks & Cleanup

After Archivist

Expected files (9 total):

README.md
background/00-dossier-overview.md
background/01-context.md
background/02-origin.md
background/03-struggle.md
background/04-pivot-point.md
background/05-renaissance.md
background/06-mechanism.md
background/07-impact.md

Common mistakes to auto-fix:

• background/00-summary.md → rename to 00-dossier-overview.md
• background/01-context-era.md → rename to 01-context.md
• background/02-origin-intended-use.md → rename to 02-origin.md
• background/03-struggle-failure.md → rename to 03-struggle.md
• background/dossier-overview.md → rename to 00-dossier-overview.md

Verification:

• All 9 files exist
• README.md is 400-600 words
• Background files cite sources

After Showrunner

Expected files (5 total in shownotes/):

shownotes/00-episode-structure.md
shownotes/01-cold-open.md
shownotes/02-narrative-arc.md
shownotes/03-playbook.md
shownotes/04-grading-criteria.md

Common mistakes to auto-fix:

• episode-structure.md at root → move to shownotes/00-episode-structure.md
• shownotes/EPISODE_STRUCTURE.md → rename to 00-episode-structure.md
• shownotes/02-chapter-outline.md → rename to 02-narrative-arc.md
• Duplicate files (e.g., both EPISODE_STRUCTURE.md and 01-episode-structure.md) → keep numbered one, delete duplicate

Verification:

• All 5 files exist in shownotes/
• Episode structure has timing information
• Grading criteria has weighted rubrics

After Podcast-writer

Expected files (9 total in transcript/):

transcript/01-cold-open.md
transcript/02-context.md
transcript/03-origin.md
transcript/04-struggle.md
transcript/05-pivot-point.md
transcript/06-renaissance.md
transcript/07-impact.md
transcript/08-playbook.md
transcript/09-grading.md

Common mistakes to auto-fix:

• transcript/02-chapter1-wrong-target.md → rename to 02-context.md
• transcript/easter-island.md → rename based on position to appropriate chapter
• Story-specific names → standardize to chapter names
• Missing files → flag as error (can't auto-fix dialogue content)

Verification:

• All 9 files exist in transcript/
• Each file has MARCUS: and ELENA: speakers
• Speaker distribution roughly balanced (45-55%)

After Editor

Expected files (1 new file):

EDITORIAL_NOTES.md

Common mistakes to auto-fix:

• 00-EDITORIAL-NOTES.md → rename to EDITORIAL_NOTES.md
• EDITORIAL_SUMMARY.md → rename to EDITORIAL_NOTES.md
• editorial-notes.md → rename to EDITORIAL_NOTES.md (uppercase)

Verification:

• EDITORIAL_NOTES.md exists at root
• Contains summary of changes made

After Speaker

Expected files (1 audio file):

transcript/<drug>-full-episode.mp3

Common mistakes to auto-fix:

• transcript/full-episode.mp3 → rename to <drug>-full-episode.mp3
• transcript/episode.mp3 → rename to <drug>-full-episode.mp3
• transcript/<drug>-<disease>-full-episode.mp3 → rename to <drug>-full-episode.mp3

Verification:

• MP3 file exists
• File size > 1MB (has actual content)

Cleanup Actions

When issues are detected, automatically:

1. Rename files with wrong names:

   mv "old-name.md" "correct-name.md"
   

2. Move files to correct locations:

   mv "episode-structure.md" "shownotes/00-episode-structure.md"
   

3. Delete duplicate files:

   rm "EPISODE_STRUCTURE.md"  # Keep the numbered version
   

4. Report cleanup to user:

   ✓ Cleaned up file structure after [Agent] completed:
     - Renamed: background/00-summary.md → 00-dossier-overview.md
     - Moved: episode-structure.md → shownotes/00-episode-structure.md
     - Deleted duplicate: EPISODE_STRUCTURE.md
   

When Cleanup Cannot Fix

If issues cannot be auto-fixed, use retry-with-feedback mechanism:

Step 1: Identify Issues

Detect problems that require agent re-work:

• Missing required files
• Empty or corrupted files
• Files with error messages instead of content
• Wrong file structure that can't be auto-renamed
• Content quality issues (no citations, missing timings, etc.)

Step 2: Re-invoke Agent Once with Feedback

Give the agent ONE chance to fix issues by re-invoking with specific feedback:

Task tool parameters:
- subagent_type: "[agent-name]"
- description: "Fix issues in [agent] output"
- prompt: "You previously generated output for [drug]-[disease] but there were issues.

           Problems found:
           - [Specific issue 1]
           - [Specific issue 2]
           - [Specific issue 3]

           Please fix these issues. Your output should include:
           - [Expected file 1]
           - [Expected file 2]
           - [Expected file 3]

           Do NOT repeat the mistakes. Follow the file structure exactly as specified
           in your instructions."

Feedback should be specific:

• ✅ "Missing required file: background/04-pivot-point.md"
• ✅ "File background/02-origin.md is empty (0 bytes)"
• ✅ "README.md is only 150 words, needs 400-600 words"
• ✅ "No sources cited in any background files"
• ❌ "Files are wrong" (too vague)

Step 3: Verify Again After Retry

After agent re-runs:

1. Check if issues are fixed
2. Run cleanup again (in case of new mistakes)
3. If STILL broken → treat as failure

Step 4: Fail After One Retry

If agent fails twice (original + one retry), give up and follow error handling:

• Create PRODUCTION_ERROR.md with specific details
• Note that agent was given feedback and failed to fix
• Commit partial work
• Report failure to user

Example error report:

## Error Description

Archivist agent failed to create required files after two attempts.

First attempt issues:
- Missing file: background/04-pivot-point.md
- File background/02-origin.md was empty

Agent was re-invoked with specific feedback.

Second attempt issues:
- Still missing file: background/04-pivot-point.md
- File now has content but no citations

After two attempts, agent could not produce correct output.

## Required Action

TODO: Human needs to manually create background/04-pivot-point.md

Retry vs. Auto-Fix Decision

Auto-fix (no retry needed):

• Wrong filename (can rename)
• Wrong location (can move)
• Duplicate files (can delete)
• Incorrect casing (can rename)

Retry with feedback:

• Missing files (agent needs to create content)
• Empty files (agent needs to write content)
• Missing critical elements (citations, timings, etc.)
• Quality issues (too short, wrong format, etc.)

Fail immediately (no retry):

• Agent crashes or returns error
• Agent refuses to run
• Downstream dependency failures (e.g., showrunner can't run if archivist failed)

Agent Configuration

All agents are defined in .claude/agents/ directory:

• archivist.md
• showrunner.md
• podcast-writer.md
• editor.md
• speaker.md
• feedback.md (used separately after episode completion)

These files contain detailed instructions for each agent's specific responsibilities.

Integration with Feedback Loop

After episode completion and user review:

1. User collects listener feedback in stories/pair-<drug>-<disease>/feedback.md
2. After 2-3 episodes, user runs feedback agent separately
3. Feedback agent updates agent instructions in .claude/agents/
4. Future episodes benefit from improved instructions

The podcast-producer skill does NOT automatically run the feedback agent. Feedback is a separate, periodic process run across multiple episodes.

Example Usage

User: "Create an episode for Aspirin and Cardiovascular Disease Prevention"

Skill Actions:

1. Parse: drug="aspirin", disease="cardiovascular-disease-prevention"
2. Create: stories/pair-aspirin-cardiovascular-disease-prevention/
3. Run archivist agent → research background
4. Run showrunner agent → create structure
5. Run podcast-writer agent → generate transcript
6. Run editor agent → editorial pass
7. Run speaker agent → generate audio
8. Commit all files to git (no push)
9. Report: "Episode complete! Audio at stories/pair-aspirin-cardiovascular-disease-prevention/transcript/aspirin-full-episode.mp3"

Success Criteria

Episode production succeeds when:

• All 5 agents complete without errors
• All expected output files exist with substantial content
• Git commit created with complete work
• User receives clear completion summary
• Episode is ready for review and publishing

Common Issues

Agent Timeout

If an agent runs too long, it may timeout. Consider:

• Running agents with longer timeout if available
• Breaking into smaller manual steps for complex episodes
• Checking if API rate limits are being hit (Gemini TTS, web search)

Missing Dependencies

If agents fail due to missing tools:

• Archivist needs: web search, academic databases
• Speaker needs: Google Gemini API key, audio processing tools
• All agents need: file system access, proper directory permissions

Git Conflicts

If git commit fails due to conflicts:

• Check if files are being modified elsewhere
• Ensure working directory is clean before starting
• Manually resolve conflicts and re-commit