Advanced Workflows

Main: "Research and write a report on remote work trends"
├── Subagent 1: Find recent statistics and studies
├── Subagent 2: Interview summaries and expert quotes
├── Subagent 3: Draft the executive summary
└── Subagent 4: Format and proofread final document

Main: "Plan our family budget for 2025"
├── Subagent 1: Analyze current spending categories
├── Subagent 2: Research cost-saving opportunities
├── Subagent 3: Build the monthly budget plan
└── Subagent 4: Create tracking spreadsheet template

Main: "Research AI safety"
├── Subagent 1: Literature review
├── Subagent 2: Current developments
├── Subagent 3: Expert opinions
└── Subagent 4: Synthesis and summary

❌ Bad: "Help with the budget"
✅ Good: "Analyze our last 3 months of spending. Identify the top 3 categories where we can realistically cut 15% or more, and explain why."

1. Enable shared topic in your main conversation
2. Spawn subagents - they automatically inherit the topic workspace
3. All subagents collaborate in the same directory
4. Results persist and are available to the main conversation

Shared Topic: "Home Purchase 2025"

Conversations:
├── "Market Researcher" (finds listings, pricing trends)
├── "Financial Planner" (budget, mortgage scenarios)
├── "Neighborhood Scout" (schools, commute, amenities)
├── "Inspector Prep" (what to look for, checklists)
└── "Document Tracker" (paperwork, deadlines)

Shared Topic: "Content Pipeline"

Pipeline:
"Research" → "Drafting" → "Editing" → "Publishing"

Research Conversation:
- Gathers information
- Stores findings in shared memory
- Saves sources to shared workspace

Drafting Conversation:
- Retrieves research from memory
- Reads source documents
- Creates draft in shared workspace

Editing Conversation:
- Reads draft
- Refines content
- Applies style guidelines

Publishing Conversation:
- Reads final draft
- Formats for publication
- Handles deployment

Shared Topic: "Code Project"

Conversations:
├── "Implementation" (primary development)
└── "Code Review" (analysis and feedback)

Implementation:
1. Writes code
2. Commits to shared workspace
3. Requests review

Code Review:
1. Reads code from workspace
2. Analyzes for issues
3. Stores feedback in memory

Implementation:
1. Retrieves feedback from memory
2. Implements improvements
3. Cycle repeats

Topic: "Home Buying 2025"

Directory Structure:
~/SAM/Home Buying 2025/
├── research/
├── finances/
├── neighborhoods/
├── viewings/
└── documents/

Week 1:
- Market Research: Average prices, trends by area
- Budget & Finances: Pre-approval estimate, monthly payment scenarios

Week 2:
- Neighborhood Comparison: Schools, commute, amenities for shortlist
- Market Research: Dig into specific listings

Week 3:
- Viewings: Log notes from each visit
- Document Checklist: Track what's needed for offer

Week 4:
- All conversations: Final comparison, make offer

Topic: "Research Paper on AI Ethics"

Conversations:
1. "Literature Review"
2. "Data Collection"
3. "Analysis"
4. "Writing"
5. "Citations"

Literature Review:
- Imports multiple PDFs
- Uses Vector RAG for semantic search
- Stores summaries in shared memory
- Tags: "methodology", "findings", "critique"

Data Collection (spawns subagents):
├── Subagent: Web research (latest developments)
├── Subagent: Expert interviews (contact info)
└── Subagent: Dataset analysis

Writing (uses all previous work):
- Retrieves summaries from memory
- Accesses imported papers
- References data collection results
- Synthesizes into coherent paper

Citations:
- Scans all references in paper
- Generates bibliography
- Verifies citation format

#!/bin/bash
# Automated code review script

# Start conversation
CONV_ID=$(curl -X POST http://localhost:8080/v1/conversations \
  -H "Content-Type: application/json" \
  -d '{"title":"Automated Review"}' \
  | jq -r '.id')

# Submit code for review
curl -X POST http://localhost:8080/api/chat/completions \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"gpt-4\",
    \"conversationId\": \"$CONV_ID\",
    \"messages\": [{
      \"role\": \"user\",
      \"content\": \"Review this PR for security and performance issues\"
    }]
  }"

# cron: 0 18 * * * /path/to/daily_summary.sh

#!/bin/bash
# Generate daily summary of project progress

curl -X POST http://localhost:8080/api/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4",
    "conversationId": "'$PROJECT_CONV_ID'",
    "messages": [{
      "role": "user",
      "content": "Summarize what I worked on today and list any open questions"
    }]
  }' > ~/daily-summary.txt

import requests

files = ['chapter1.md', 'chapter2.md', 'chapter3.md']
endpoint = 'http://localhost:8080/api/chat/completions'

for file in files:
    response = requests.post(endpoint, json={
        'model': 'gpt-4',
        'conversationId': conv_id,
        'messages': [{
            'role': 'user',
            'content': f'Analyze {file} for code quality issues'
        }]
    })
    print(f"Results for {file}:", response.json())

Small Tasks       → Default (low scaling)
Long Conversations → Extended (medium scaling)
Document Analysis  → Universal (high scaling, default)
Enterprise Docs    → Mega (enterprise scaling)

When context fills:
1. Clear less important messages
2. Summarize earlier parts
3. Store summaries in memory
4. Continue with clean context

Every few weeks:
1. Review memory statistics
2. Clear low-importance memories (< 0.4)
3. Remove duplicate entries
4. Archive completed project memories

❌ Don't store everything
✅ Store decisions, requirements, and critical information

❌ Don't duplicate across conversations
✅ Use shared topics for related work

❌ Don't use high similarity thresholds for documents
✅ Lower threshold for document search (0.15-0.25)

Quick Q&A          → GPT-3.5-turbo (fast, cheap)
Complex Logic      → GPT-4 (best reasoning)
Creative Writing   → Claude 3.5 Sonnet (creative)
Code Generation    → GitHub Copilot GPT-4 (code-optimized)
Privacy-Critical   → Local MLX/GGUF models (offline)

1. Use cheaper models (GPT-3.5-turbo) for initial drafts
2. Refine with expensive models (GPT-4) when needed
3. Use local models for sensitive or privacy-critical data
4. Enable streaming for better user experience

Step 1: Web Research
Tool: web_operations (research)
Result: Stores findings in memory

Step 2: Memory Retrieval
Tool: memory_operations (search_memory)
Result: Retrieves relevant research

Step 3: Document Creation
Tool: document_operations (create)
Result: Creates final document with citations

Session 1: "build"
- run_command: "make clean"
- run_command: "make build"
- get_output: Check for errors

Session 2: "test"
- run_command: "pytest tests/"
- get_output: Verify all passed

Session 3: "deploy"
- run_command: "docker build ."
- run_command: "docker push ..."

Main conversation maintains multiple sessions:
- "research": Gathering sources and data
- "writing": Drafting sections
- "review": Fact-checking and editing
- "output": Formatting and export

Switch between sessions as needed

Instead of:
- create_file (10 times)

Use:
- multi_replace_string (one operation)
- Apply templates efficiently

Semantic: Find by meaning
"notes about our budget constraints"

Regex: Find by pattern
"Total:.*\$[0-9]+"

Glob: Find by filename
"**/budget-*.md"

1. Create shared topic: "AI in Healthcare Report"
2. Create four conversations:
   - "Literature Review" (Scholar personality, GPT-4, working dir: sources/)
   - "Data Analysis" (Professional personality, Claude 3.5, working dir: data/)
   - "Writing" (Creative personality, GPT-4, working dir: drafts/)
   - "Fact-Checking" (Scholar personality, GPT-4, working dir: sources/)
3. Enable shared topic in all conversations

You: Find and summarize 10 key papers on AI diagnostics published since 2021
SAM: [Searches web for papers]
     Summarized 10 papers - saved to sources/literature-review.md
     Key themes: accuracy improvements, FDA approval challenges, bias concerns

You: Read the literature review and identify the key statistics to highlight
SAM: [Reads sources/literature-review.md from shared workspace]
     Found 8 compelling statistics - saved to data/key-stats.md
     Strongest: "AI matched radiologist accuracy in 94% of cases (Smith et al, 2023)"

You: Write the executive summary using our research and key stats
SAM: [Reads literature-review.md and key-stats.md]
     Draft saved to drafts/executive-summary.md
     ~800 words, covers all major themes

You: Check every statistic in the executive summary against our sources
SAM: [Reads drafts/executive-summary.md and sources/literature-review.md]
     All 6 statistics verified. One citation needed a correction - fixed in draft.