synthesize（synthesize 搜索）

Create and incrementally build synthesis documents from multiple sources.

Output Location

All synthesis documents go in: my-vault/06 Knowledge Base/Synthesis/[Topic]/

Structure:

my-vault/06 Knowledge Base/Synthesis/
└── ADHD/
    ├── ADHD Synthesis.md              # Main synthesis
    ├── ADHD Sources.md                # Sources table + processing log
    └── ADHD Medications Synthesis.md  # Sub-topic (if created)

Queue Files

Queues live in: .claude/skills/synthesize/references/queues/[topic].json

{
  "name": "ADHD",
  "topic_folder": "ADHD",
  "batch_size": 3,
  "max_discovered_per_source": 3,
  "watch_patterns": [
    "my-vault/06 Knowledge Base/Capture/Videos/**/*ADHD*.md",
    "my-vault/01 Inbox/*adhd*.md",
    "my-vault/04 Personal/Health/*ADHD*.md"
  ],
  "sources": [
    {"path": "my-vault/path/to/note.md", "status": "pending", "type": "local"},
    {"path": "my-vault/path/to/video.md", "status": "processed", "type": "local"},
    {"url": "https://example.com/study", "status": "pending", "type": "web",
     "discovered_from": "src-2", "context": "Key study on medication timing"},
    {"search": "Barkley 2023 executive function", "status": "pending", "type": "search",
     "discovered_from": "src-2", "context": "Referenced as foundational research"}
  ],
  "last_run": "2026-01-14T10:00:00Z"
}

Source types:

• local - Path to existing note in my-vault
• web - URL to fetch and process
• search - Search query to find the source (use WebSearch, then fetch best result)

Source status values: pending, processed, skipped

watch_patterns: Glob patterns for /synthesize update to scan. Paths matching these patterns get added to the queue automatically (if not already present).

Commands

• /synthesize [topic] - Continue processing the named synthesis
• /synthesize list - Show all active synthesis queues and progress
• /synthesize new [topic] - Create a new synthesis queue
• /synthesize add [topic] [path-or-glob] - Add sources to a queue
• /synthesize update [topic] - Scan watch_patterns for new sources and add to queue

Architecture

This skill uses a coordinator + subagent pattern for efficient processing:

• Coordinator (Sonnet/Opus): Manages state, spawns extractors, merges findings into synthesis
• Extractors (Haiku subagents): Each reads one source, extracts structured findings in parallel

This allows processing 6-10 sources per batch instead of 2-3, with ~3x throughput improvement.

Workflow

Starting a Session

1. Load queue file for the topic
2. Read the main synthesis doc (or create from template if first run)
3. Read the sources file
4. Identify next N pending sources (N = 6-10 for parallel extraction)

Phase 1: Parallel Extraction (Haiku Subagents)

Spawn 6-10 Haiku extractors in parallel using the Task tool:

Task tool call:
- model: haiku
- subagent_type: general-purpose
- prompt: (see Extractor Prompt Template below)

Extractor Prompt Template:

Read this file and extract [TOPIC]-relevant findings in a structured format.

FILE: [full path to source]

Return your findings as a simple list in this format:
CATEGORY: [Category Name]
- Finding 1 (be specific, include numbers/names if mentioned)
- Finding 2
...

Only include substantive findings relevant to understanding [TOPIC].
Skip metadata, timestamps, and filler.
If the source has little new information, just return "MINIMAL_CONTENT".

For web sources, modify prompt to use WebFetch first:

1. Use WebFetch to retrieve: [URL]
2. Extract [TOPIC]-relevant findings from the content...

For search sources, modify prompt to search first:

1. Use WebSearch to find: [search query]
2. Use WebFetch on the best result
3. Extract [TOPIC]-relevant findings...

Phase 2: Merge Findings (Coordinator)

Once all extractors return:

1. Collect all extraction results
2. Review for:

- New findings to add to existing categories - Findings that warrant new categories - Contradictions with existing content → add to Open Questions - Dense categories → add to Suggested Sub-Topics

1. Update synthesis doc with citations (^[src-N])
2. Add sources to Sources table
3. Mark sources as processed in queue

Phase 3: Discovery (Coordinator)

While reviewing extractions, watch for references worth following up:

• Studies or papers cited
• Articles or books mentioned
• Other videos or talks referenced
• Experts or researchers named

Limit: 3 discovered sources per source processed.

Discovery

While processing each source, watch for references worth following up:

• Studies or papers cited
• Articles or books mentioned
• Other videos or talks referenced
• Experts or researchers named (for potential lookup)

Limit: 3 discovered sources per source processed.

For each valuable reference:

1. Check if already in queue (deduplicate by URL, title, or path)
2. Add to queue with:

- type: web if URL available, search if just a name/title - discovered_from: source ID where you found it - context: why it's relevant (1 sentence) - status: pending

Prioritize discoveries that:

• Are directly cited as evidence for claims
• Come from credible sources (peer-reviewed, recognized experts)
• Fill gaps in the current synthesis

Skip discovering:

• Tangential references not central to the topic
• Sources that would require paid access (note in processing log if significant)
• Generic resources (e.g., "see Wikipedia for more")

Ending a Session

1. Write updated synthesis doc
2. Write updated sources file
3. Write updated queue file with new last_run timestamp
4. Add entry to Processing Log (include discovery count)
5. Report: ` Processed X items. Discovered Y new sources. Z items remaining. Run /synthesize [topic] to continue. `

If all sources processed:

Synthesis complete. X total sources processed.
Review Suggested Sub-Topics for potential expansion.

Updating a Synthesis (/synthesize update [topic])

Use this to find new sources that match watch_patterns without reprocessing old ones.

1. Load queue file for the topic
2. Collect all paths from watch_patterns using Glob
3. Filter out any paths already in sources (regardless of status)
4. Add new matches to sources with:

- type: local - status: pending

1. Update queue file
2. Report: ` Found X new sources matching watch_patterns. Queue now has Y pending items. Run /synthesize [topic] to process. `

If no new sources found:

No new sources found. Synthesis is up to date with watch_patterns.

Tip: Run /synthesize update [topic] periodically after adding new notes or video summaries to keep the synthesis current.

Synthesis Document Template

Create at: my-vault/06 Knowledge Base/Synthesis/[Topic]/[Topic] Synthesis.md

---
class: Synthesis
topic: [Topic Name]
created: [Date]
lastUpdated: [Date]
status: in-progress
---

## Overview

[2-3 sentence summary of the topic - updated as understanding deepens]

## Key Findings

### [Category]

- Finding with specific details ^[src-1]
- Another finding that multiple sources support ^[src-1] ^[src-3]

<!-- Add categories as needed based on source content -->

## Open Questions

<!-- Contradictions, uncertainties, areas needing human judgment -->

- **[Topic]**: Source A says X ^[src-1], but Source B says Y ^[src-2].

## Suggested Sub-Topics

<!-- Categories that could warrant their own synthesis -->

- **[Sub-topic]** - X sources touched on this; may warrant dedicated synthesis

---
*Sources and processing log: [[Topic Sources]]*

Sources Document Template

Create at: my-vault/06 Knowledge Base/Synthesis/[Topic]/[Topic] Sources.md

---
class: Note
parent: "[[Topic Synthesis]]"
---

## Sources

| ID | Title | Date | Type | Discovered | Path/URL |
|----|-------|------|------|------------|----------|
| src-1 | Source Title | 2025-03 | Video | - | [[path/to/note]] |
| src-2 | Study Name | 2024-11 | Web | from src-1 | [link](url) |

## Processing Log

- **2026-01-14**: Initial synthesis. Processed 3 items. Created categories for X, Y, Z. Discovered 2 new sources.
- **2026-01-15**: Processed 3 items. Flagged contradiction on [topic]. Discovered 1 new source.

Length Discipline

Core content (Overview through Suggested Sub-Topics) must stay under 300 lines.

Guidelines:

• Overview: 5 lines max
• Categories: 8 max in main doc
• Points per category: 15 max
• Open Questions: 10 max

When a category exceeds ~40-50 lines or has 5+ sources going deep:

1. Summarize tightly in the main doc (5-10 key points)
2. Add to Suggested Sub-Topics
3. Sub-topic syntheses get their own doc in the same folder, linking back to parent

Handling Contradictions

When new information conflicts with existing content:

1. Do not silently replace the existing information
2. Move both claims to Open Questions with citations: - **Medication timing**: Take with food for reduced side effects ^[src-2], but empty stomach for faster absorption ^[src-5]. May depend on specific medication.
3. If a clear resolution exists (newer research, more credible source, broader consensus), note it: - **Resolved**: Earlier sources suggested X, but consensus now supports Y ^[src-8] ^[src-9] ^[src-10]

Citation Format

Use ^[source-id] inline, where source-id matches the ID column in the Sources table.

• Single source: Finding here ^[src-1]
• Multiple sources: Well-supported finding ^[src-1] ^[src-3] ^[src-7]
• Conflicting sources: Note in Open Questions section

Adding Sources to Queue

Manual Addition

/synthesize add adhd my-vault/path/to/specific/note.md

Glob Pattern

/synthesize add adhd "my-vault/06 Knowledge Base/Capture/Videos/**/*ADHD*.md"

By Tag (search and add)

When adding, search for relevant notes:

• Notes with topic in filename
• Notes with relevant tags
• Video summaries from relevant channels

Deduplicate against existing sources in queue before adding.

Creating Sub-Topic Syntheses

When a sub-topic warrants its own document:

1. Create new queue: /synthesize new "ADHD Medications"
2. Set topic_folder to match parent: "ADHD" (keeps files together)
3. Add relevant sources (can include already-processed sources from parent for deeper dive)
4. In parent synthesis, add link: See also: [[ADHD Medications Synthesis]]
5. In sub-topic synthesis Overview, note parent: Deep-dive from [[ADHD Synthesis]]