Skip to content

handover.md

Latest commit

 

History

History
418 lines (321 loc) · 17.4 KB

handover.md

File metadata and controls

418 lines (321 loc) · 17.4 KB
Raw

COMPREHENSIVE HANDOVER - Integrated Content Operating System

Last Updated: 2025-12-31 Sessions: System Integration → Environment Setup → Multi-LLM Setup

LATEST SESSION: Multi-LLM Parallel Writing Setup (2025-12-31)

What Was Accomplished

1. Created GEMINI.md for Gemini CLI ✅

Created parallel context file for Gemini CLI at GEMINI.md:

  • Same skills, pipelines, voice standards as CLAUDE.md
  • Added coordination section for working alongside Claude Code
  • Highlighted Gemini-specific strengths (long context, research, images)
  • Added session management and handover protocols

2. Multi-LLM Coordination System ✅

Model Context File Strengths
Claude Code CLAUDE.md Tool use, code editing, YouTube scripts
Gemini CLI GEMINI.md Research, long docs, images, free tier
Codex CLI AGENTS.md Script execution, structured content, batch generation

Switching Model: On-demand - use whichever is available/appropriate

4. Created AGENTS.md for Codex CLI ✅

Converted Codex from code-only to content writer:

  • Same skills, pipelines, voice standards
  • Explicit instruction that it's a WRITER not coder
  • Highlighted its strengths: script execution, structured output, batch processing
  • Coordination with Claude and Gemini

3. Created Sync Script ✅

scripts/sync_llm_contexts.py - Keeps CLAUDE.md and GEMINI.md in sync:

python scripts/sync_llm_contexts.py --check          # Check sync status
python scripts/sync_llm_contexts.py --show-diff      # Show differences
python scripts/sync_llm_contexts.py --update-gemini  # Sync from CLAUDE.md

Sections that sync: Skills, Pipelines, Directory Structure, Voice Standards, API Keys Sections that DON'T sync: Model-specific instructions, coordination, strengths

MULTI-LLM COORDINATION (3 MODELS)

How It Works

  1. Claude Code, Gemini CLI, and Codex CLI can all be used in parallel
  2. Check this handover.md at the start of any session
  3. Update handover.md when switching or stopping mid-task
  4. Don't overwrite files another model is actively editing

The Three Writers

Model Context File Cost Strengths
Claude Code CLAUDE.md $20/mo Pro Voice accuracy, tool use, YouTube scripts
Gemini CLI GEMINI.md FREE / $20 Pro Long context, research, images
Codex CLI AGENTS.md $20/mo (ChatGPT Pro) Script execution, structured content, batch

Task Routing Suggestions

Task Best Model Why
YouTube scripts (Hinglish) Claude Voice training
Research synthesis Gemini 1M token context, grounding
Newsletter drafts Gemini Long document handling
Image generation Gemini Native API
Structured editorials Codex Analytical, tables, data
Batch tweet generation Codex Efficient bulk output
Script execution Codex Code execution ability
Code editing Claude Better tool use
Twitter threads Any All capable
Trial analysis Codex or Claude Both trained

Handover Format (When Switching)

## Active Work: [Date/Time]
**Model:** [Claude/Gemini]

### Completed
- [x] Task 1

### In Progress
- [ ] Task 2 - stopped at: [description]
  - File: path/to/file
  - Next step: [what to do]

PREVIOUS SESSION: Environment & Security Setup (2025-12-29)

What Was Accomplished

1. Environment Variables Configured ✅

Created .env with 35 configuration values covering all services:

Category Keys Added
LLM Providers Anthropic (2), OpenAI, OpenRouter (2), Gemini (2), xAI, Groq, Perplexity, Cohere
Research NCBI/PubMed, AstraDB (4 params), Tavily
Data/Automation Airtable (2), Apify, Bright Data, Supadata
YouTube/Video YouTube API (2), FAL.AI
Development GitHub, LangSmith (+ tracing config)
Email Gmail SMTP (user + app password)
Browser Automation ChatGPT email, Gemini email

2. Security Infrastructure Created ✅

  • Created .gitignore - Protects .env and other secrets from git commits
  • File permissions set to owner-only (-rw-------)

3. Security Audit Completed ✅

Ran multi-agent security scan (4 parallel haiku agents) to find exposed secrets.

Findings Fixed:

File Issue Resolution
workflows/BUILD-THIS-PROJECT.md Real OpenRouter, Slack, NCBI keys exposed Replaced with placeholders
workflows/MEDICAL-CONTENT-ENGINE-SPEC.md Real API keys + hardcoded email Replaced with ${VAR_NAME} references
Code examples in docs Hardcoded email addresses Changed to os.getenv() pattern

Safe (No Action Needed):

  • All Python/JS code already uses os.getenv() correctly
  • .mcp.json uses ${NCBI_API_KEY} placeholder correctly
  • .env.example contains only template placeholders

4. User Decisions Made

  • Slack integration: Will NOT be used (to be changed later)
  • API key method: User provided keys via conversation (acceptable for dev keys)
  • Gmail SMTP: App password added for email workflows

SOURCE OF TRUTH

SINGLE LOCATION: /Users/shaileshsingh/integrated cowriting system/

This directory is now the ONLY place for all content creation, research, and writing tools. Works with Antigravity, VS Code, Cursor, and Claude Code.

Master Instructions: Read CLAUDE.md first - contains complete system documentation.

WHAT WAS ACCOMPLISHED THIS SESSION

1. Full System Integration Complete ✅

Consolidated 40+ scattered directories into ONE unified system:

Component Source Destination Status
135 Scientific Skills Downloads/claude-scientific-skills-main.zip skills/scientific/ ✅ Installed
33 Cardiology Skills Downloads/*.skill + cowriting system/.claude/skills/ skills/cardiology/ ✅ Installed
Twitter Pipeline new twitter content system/ pipelines/twitter-content/ ✅ Copied
YouTube Research cowriting system/research-engine/ research-engine/ ✅ Copied
RAG Pipeline cowriting system/rag-pipeline/ rag-pipeline/ ✅ Copied
Antigravity Workflows fetch journal articles/workflows/ workflows/ ✅ Copied
Context Profiles cowriting system/context-profiles/ context-profiles/ ✅ Copied
PubMed MCP Configuration created .mcp.json ✅ Configured

2. Created Master CLAUDE.md ✅

Comprehensive documentation covering:

  • System architecture diagram
  • All 3 pipelines (Twitter, YouTube, Journal)
  • PubMed MCP integration (MANDATORY)
  • AstraDB RAG integration (MANDATORY)
  • All 168 skills documented
  • Content workflows
  • API keys required
  • Anti-AI guidelines
  • Quick start guide

3. Directory Structure Created ✅

integrated cowriting system/
├── CLAUDE.md                    # Master instructions
├── .mcp.json                    # PubMed MCP config
├── .env.example                 # API keys template
├── requirements.txt             # Dependencies
├── skills/
│   ├── scientific/              # 135 skills (alphafold, biopython, etc.)
│   └── cardiology/              # 33 skills (youtube-script-master, etc.)
├── pipelines/
│   └── twitter-content/         # Tweet fetching & generation
├── research-engine/             # YouTube research (35+ channels)
├── rag-pipeline/                # AstraDB RAG (8 techniques)
├── workflows/                   # 4 Antigravity/n8n workflows
└── context-profiles/            # Voice DNA, ICP, business profile

USER DECISIONS MADE (CRITICAL)

What User WANTS:

  1. ✅ PubMed MCP - MANDATORY for all content (no citation management needed separately)
  2. ✅ AstraDB RAG - MANDATORY for guidelines/textbooks
  3. ✅ Tweet Fetching System - via Apify (8 medical influencers)
  4. ✅ Journal Fetching - via Antigravity workflows (20+ journals)
  5. ✅ YouTube Research System - 35+ channels, 8 narratives tracked
  6. ✅ ALL 135 Scientific Skills from GitHub repo
  7. ✅ ALL Cardiology Skills (33 total)
  8. ✅ This directory as SINGLE SOURCE OF TRUTH
  9. ✅ Works with Antigravity, VS Code, Cursor

What User DOES NOT WANT:

  1. ❌ DeepResearch Draft Auditor - NOT needed
  2. ❌ Medical Ghostwriter's memory/citation features - NOT needed
  3. ❌ Separate hallucination detection - NOT needed (PubMed provides real citations)

SYSTEM ARCHITECTURE

┌─────────────────────────────────────────────────────────────────────────────┐
│                    INTEGRATED CONTENT OPERATING SYSTEM                       │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                              │
│  ┌──────────────────┐  ┌──────────────────┐  ┌──────────────────┐           │
│  │  TWITTER SYSTEM  │  │  YOUTUBE SYSTEM  │  │  JOURNAL SYSTEM  │           │
│  │  Apify Scraping  │  │  35+ Channels    │  │  20+ Journals    │           │
│  │  8 Influencers   │  │  8 Narratives    │  │  B2C/B2B Triage  │           │
│  └────────┬─────────┘  └────────┬─────────┘  └────────┬─────────┘           │
│           │                     │                     │                      │
│           └─────────────────────┼─────────────────────┘                      │
│                                 │                                            │
│                    ┌────────────▼────────────┐                               │
│                    │     RESEARCH LAYER      │                               │
│                    ├─────────────────────────┤                               │
│                    │  PubMed MCP (MANDATORY) │                               │
│                    │  AstraDB RAG (MANDATORY)│                               │
│                    └────────────┬────────────┘                               │
│                                 │                                            │
│                    ┌────────────▼────────────┐                               │
│                    │     168 SKILLS          │                               │
│                    │  135 Scientific         │                               │
│                    │  33 Cardiology          │                               │
│                    └─────────────────────────┘                               │
│                                                                              │
└─────────────────────────────────────────────────────────────────────────────┘

VOICE STANDARDS (Important)

Content Type Voice Model Language
YouTube Peter Attia speaking Hinglish 70% Hindi / 30% English
Twitter/Writing Eric Topol Ground Truths English
B2B (Doctors) JACC editorial style English

WHAT IS PENDING

1. Environment Setup ✅ COMPLETE

  • Created .env with 35 configuration values
  • Created .gitignore to protect secrets
  • Security audit completed - all exposed secrets fixed
  • Gmail SMTP configured with app password

2. Test Pipelines

  • Test Twitter pipeline: python pipelines/twitter-content/generate.py "test question"
  • Test YouTube research: python research-engine/run_pipeline.py --quick
  • Test RAG pipeline: Import and run knowledge_pipeline
  • Test Antigravity workflows in n8n

3. Potential Cleanup (User Decision Required)

  • Delete old redundant directories (40+ scattered across system)
  • User to confirm which old directories can be deleted

4. Refinements (Future)

  • Refine skills as needed
  • Add custom slash commands
  • Set up daily automation schedules
  • Replace Slack integration with alternative (user decision pending)

KEY FILES TO READ

File Purpose
CLAUDE.md READ FIRST - Complete system documentation
.env All 35 API keys and configuration (PROTECTED - in .gitignore)
.gitignore Protects secrets from git commits
.mcp.json PubMed MCP server configuration
.env.example Template for required API keys
workflows/MEDICAL-CONTENT-ENGINE-SPEC.md Journal fetching spec
workflows/PROJECT-SPEC-TWITTER-CONTENT-SYSTEM.md Twitter system spec
context-profiles/voice-dna.md Voice guidelines

ORIGINAL DIRECTORIES (For Reference)

These were the source directories for the integration:

Directory What Was Taken Status
/Users/shaileshsingh/cowriting system/ Research engine, RAG, skills, context profiles Integrated ✅
/Users/shaileshsingh/new twitter content system/ Twitter pipeline Integrated ✅
/Users/shaileshsingh/fetch journal articles/ Antigravity workflows, specs Integrated ✅
/Users/shaileshsingh/Downloads/claude-scientific-skills-main.zip 135 scientific skills Extracted ✅
/Users/shaileshsingh/Downloads/*.skill 8 cardiology skill files Copied ✅
/Users/shaileshsingh/pubmed-mcp-server/ MCP server (referenced in .mcp.json) Configured ✅

PIPELINES SUMMARY

1. Twitter Content Pipeline

Location: pipelines/twitter-content/ Flow: HARVEST → RESEARCH → SYNTHESIZE → WRITE Files: harvester.py, researcher.py, synthesizer.py, writer.py, pipeline.py Usage:

python pipelines/twitter-content/generate.py "Your question"
python pipelines/twitter-content/main.py harvest

2. YouTube Research Pipeline

Location: research-engine/ Components: Channel scraper, comment scraper, demand signals, narrative monitor, gap finder, view predictor, calendar generator Usage:

python research-engine/run_pipeline.py --quick
python research-engine/analyzer/narrative_monitor.py --debunk

3. Journal Fetching Pipeline

Location: workflows/ (Antigravity/n8n JSON files) Workflows:

  • unified-twitter-content-pipeline.json
  • medical-content-engine.json
  • medical-content-engine-v2.json
  • medical-journal-digest.json

SKILLS INVENTORY

Scientific Skills (135) - skills/scientific/

Databases, bioinformatics, data science, lab tools including:

  • alphafold-database, biopython, biorxiv-database, chembl-database
  • clinicaltrials-database, clinvar-database, pubmed-database
  • deepchem, scanpy, pytorch, scikit-learn, and 125+ more

Cardiology Skills (33) - skills/cardiology/

  • YouTube: youtube-script-master, youtube-script-hinglish, debunk-script-writer, hook-generator
  • Writing: cardiology-editorial, cardiology-newsletter-writer, cardiology-trial-editorial, x-post-creator-skill
  • Research: pubmed-database, clinicaltrials-database, literature-review, citation-management
  • Quality: authentic-voice, content-reflection, scientific-critical-thinking

HOW TO CONTINUE NEXT SESSION

1. Open: /Users/shaileshsingh/integrated cowriting system/
2. Read: CLAUDE.md (master instructions)
3. Read: handover.md (this file)
4. Set up: .env file with API keys
5. Test: Run pipelines to verify everything works
6. Optionally: Confirm deletion of old directories

CRITICAL REMINDERS

  1. PubMed MCP is MANDATORY - Nothing goes out without PubMed verification
  2. AstraDB RAG is MANDATORY - All content must reference knowledge base
  3. This is the SINGLE SOURCE OF TRUTH - Don't create content elsewhere
  4. Works with: Antigravity, VS Code, Cursor, Claude Code
  5. 168 Skills available - 135 scientific + 33 cardiology

CONTEXT WINDOW PROTOCOL

When context < 20%:

  1. STOP current work
  2. Create comprehensive handover.md
  3. Inform user
  4. Wait for new session

SESSION HISTORY

Date Session Key Accomplishments
2025-12-29 System Integration Consolidated 40+ directories, created CLAUDE.md, installed 168 skills
2025-12-29 Environment & Security Created .env (35 keys), .gitignore, security audit, fixed exposed secrets
2025-12-31 Multi-LLM Setup Created GEMINI.md + AGENTS.md, 3-model parallel system, sync script

End of Handover Last Updated: 2025-12-31 Latest Session: Multi-LLM Parallel Writing Setup Next: Test Gemini CLI with GEMINI.md, test pipelines, explore parallel writing workflows