synthesize

Science grounding: far transfer (Tempel & Frings 2024), generation effect (McCurdy et al. 2020), elaborative interrogation, desirable difficulty (Bjork & Bjork 2020). </Purpose>

<Use_When>

• User wants to understand what the vault collectively says about a concept
• User is writing a Thought note and wants to pressure-test an idea
• User is preparing a talk, essay, or decision and needs cross-domain ammunition
• User says "synthesize X", "what do I know about X", "connect ideas about X"
• User wants to find contradictions or gaps in their own thinking </Use_When>

<Do_Not_Use_When>

• User wants to process a single source note — use /process instead
• User wants spaced repetition review — use /recall instead
• User has a specific paper to analyze — use /paper instead
• Topic is too narrow to generate cross-domain connections (single paper, single event) </Do_Not_Use_When>

Stage 1: PARSE THE TOPIC

Extract the core concept(s) from the user's input. Identify:

• Primary term: the exact word(s) to search (e.g., "attention", "loss aversion")
• Synonyms / adjacent: related terms likely to appear in notes (e.g., "focus", "salience", "bias")
• Domain anchors: which domains likely hold relevant notes (ML, psychology, design, startup...)

If the topic is ambiguous, briefly clarify before searching (one question max).

Stage 2: PARALLEL SEARCH — run ALL of these simultaneously

MCP search (preferred):

search_notes(query="PRIMARY TERM", limit=30)
search_notes(query="SYNONYM 1", limit=15)
search_notes(query="SYNONYM 2", limit=15)

Grep fallback (if MCP unavailable):

Grep(pattern="PRIMARY TERM", path="notes/", glob="*.md", output_mode="files_with_matches")
Grep(pattern="#TAG-NAME", path="notes/", glob="*.md", output_mode="files_with_matches")
Grep(pattern="\[\[.*PRIMARY TERM.*\]\]", path="notes/", glob="*.md", output_mode="files_with_matches")

Also search:

• Thought notes referencing this topic: Grep(pattern="PRIMARY TERM", path="notes/thoughts/", glob="*.md")
• Cross-domain appearances: search each domain folder separately to surface unexpected hits
• Existing Term notes: Grep(pattern="type: term", path="notes/", glob="*.md") then filter

Collect all matching file paths. Deduplicate. Aim for 15–25 notes across diverse domains.

Stage 3: READ THE NOTES

Read up to 20 most relevant notes using read_multiple_notes or individual read_note calls.

Prioritize:

1. Thought notes (user's own synthesis — highest signal)
2. Term/Note notes (distilled concepts)
3. Paper/book/post notes with substantial content
4. Notes from UNEXPECTED domains (highest cross-domain value)

While reading, track:

• Key claims and insights
• Domain of each note
• Wikilinks that connect to the topic
• Any explicit contradictions or tensions
• Claims that could conflict with each other

Stage 4: PRESENT THE SYNTHESIS

Output this structure (adapt headings as needed, cut any section with nothing real to say):

What the vault says about [TOPIC]

Key insights

[3–6 bullets. Each = one non-obvious claim distilled from multiple notes. Lead with the sharpest insight. No filler.]

Cross-domain appearances

[Table or bullet list: Domain → how the concept appears there. Include only non-obvious mappings — skip domains where the concept is trivially expected.]

Agreements

[What multiple notes converge on — the vault's "consensus". 2–4 bullets max.]

Tensions & contradictions

[Where notes disagree — including tensions the user may not have noticed. THIS IS THE MOST VALUABLE SECTION. If two notes from different domains make opposing claims, surface it explicitly. Quote the note titles.]

Gaps

[What's missing: questions the vault raises but doesn't answer, adjacent concepts not yet captured, domains that probably have something to say but no notes exist]

Stage 5: CHALLENGE QUESTIONS — do NOT skip, do NOT soften

Ask 3–4 questions. Make them uncomfortable. Reference the user's OWN notes by name.

Rules for good challenge questions:

• Anchor in specific vault evidence: "Your note [[(Thought) X]] claims Y, but [[(Paper) Z]] shows the opposite — which do you actually believe?"
• Force abstraction: "These 4 examples from different domains all share [pattern]. Is there a general principle here that deserves its own Term note?"
• Force disagreement: "What would someone who thinks [vault consensus claim] is WRONG say? Do you have any evidence for their view?"
• Force transfer: "You have strong notes on [concept A] and [concept B]. How does [concept A] change your understanding of [concept B]?"

Wait for user responses before proceeding to Stage 6.

The vault should not just confirm what you know. If the synthesis has no tension and no surprises, say so explicitly: "I found strong consensus but no real tension — either the topic is genuinely settled, or the vault has a blind spot here."

Stage 6: HELP CREATE OUTPUT (only if something emerged)

If a new Thought emerged from the conversation:

Help the user create a Thought note. Do NOT write it for them — Thoughts are where human understanding lives. Instead:

1. Ask: "What's the single sentence that captures the insight?"
2. Offer to create the note structure, leaving the content for them to fill
3. Generate the frontmatter + skeleton only

Thought note template:

---
id: YYYYMMDDHHMMSS
created_date: YYYY-MM-DD
updated_date: YYYY-MM-DD
type: thought
processing_status: processed
---

# (Thought) [Title]
- **🏷️Tags** : #thought #[topic-tag] #MM-YYYY

[ ](#anki-card)

## Notes
[User fills this]

## Links
- [[(Type) Source Note 1]]
- [[(Type) Source Note 2]]

If a new Term should be extracted:

Only propose this if the concept appeared across multiple notes but has no dedicated Term note. Create the Term note with:

---
id: YYYYMMDDHHMMSS
created_date: YYYY-MM-DD
updated_date: YYYY-MM-DD
type: term
processing_status: processed
---

# (Term) [Concept Name]
- **🏷️Tags** : #term #all-anki #[topic-tag] #MM-YYYY

[ ](#anki-card)

## Notes
[Distilled definition — Feynman-style: intuition first, formalism second]

- [Key insight 1]
- [Key insight 2]
- **Example**: [concrete example]
- **Cross-domain**: [where else this appears]

## Links
- [[(Type) Source 1]]

Wikilink rule (always):

Every internal reference must use the full filename stem including (Type) prefix:

• Correct: [[(Thought) My insight note]]
• Wrong: [[My insight note]]

If unsure of the exact filename, search before linking.

<Escalation_And_Stop_Conditions>

• Too few notes (< 3 relevant): Say so — "The vault is thin on this topic. Suggest reading X or running /paper-discover to build coverage."
• Topic too broad: Ask user to narrow. One clarifying question, not a list.
• Only one domain represented: Flag it — "All your notes on this come from ML. This might be a blind spot. Want me to /paper-discover adjacent psychology or design literature?"
• No tension found: Report honestly. Don't manufacture fake contradictions.
• User disagrees with a challenge question: Engage — don't retreat. The job is productive friction.
• User wants to skip challenge questions: Gently push back once. "The questions are the point — skip them and you're just reading an index." Then respect their choice. </Escalation_And_Stop_Conditions>

$ARGUMENTS