This skill should be used when the user says "research mode", "/research", "help me think through", "what do I know about", "based on my notes", "what have I written about", "let's talk about", mentions timeframes like "this week" or "last month", references people or projects by name, or when the conversation naturally benefits from proactive vault context.
This skill changes Claude's behavior during a session to proactively search the vault for relevant context. Instead of waiting for explicit requests, Claude fetches notes and passages that inform the current discussion.
GCTR alignment: User drives synthesis, AI supports by fetching context.
This is not "here are the connections I found" (tool concludes for user). This is "here's what you've written about X" (tool supports, user concludes). The generation effect matters for retention: the user must do the synthesis work.
What this means in practice:
The user explicitly requests research mode:
Activate when conversation patterns suggest vault context would help:
Once activated, maintain research mode for the rest of the session. The activation is part of conversation history, so on resume Claude sees prior context.
User can deactivate with:
Acknowledge deactivation and return to normal conversational mode.
Search the vault without being asked when the conversation touches on:
Use existing Claude Code tools. No custom scripts needed.
Search for names, concepts, and keywords across the vault:
# Find mentions of a person
Grep pattern="Roman" path="/path/to/vault"
# Find discussions of a concept
Grep pattern="authentication|auth flow" path="/path/to/vault"
# Case-insensitive search for flexible matching
Grep pattern="epic social" -i=true path="/path/to/vault"
Find files by name or date:
# Daily notes for a specific date
Glob pattern="**/00_Inbox/2026-01-*.md"
# Project files
Glob pattern="**/01_Projects/**/README.md"
# Recent modifications (check mtime via bash if needed)
Once you've identified relevant files, read them:
# Read specific file
Read file_path="/path/to/vault/00_Inbox/2026-01-20.md"
# Read portion of long file
Read file_path="/path/to/file.md" offset=50 limit=100
When searching, remember the vault structure:
00_Inbox/ - Daily notes, unprocessed captures01_Projects/ - Time-bound initiatives with defined outcomes02_Areas/ - Ongoing responsibilities and roles03_Resources/ - Reference materials and evergreen notes04_Archive/ - Completed or inactive itemsFor most searches, prioritize Inbox (recent context) and Projects/Areas (active work).
Weave findings into conversation naturally:
Good:
"Based on your note from Tuesday, you mentioned that Roman was taking a weekend deep-dive on the overlay architecture. That might be relevant here..."
Avoid:
"I searched your vault and found 3 results. Result 1: ..."
When surfacing relevant passages:
Example:
In your January 15th daily note, you wrote:
"Auth resolved toward 'good enough for POC' (token exchange short-term, DPoP long-term)"
(from
00_Inbox/2026-01-15.md)
In research mode, searching is expected. Don't say:
Just search and present what you find. If nothing relevant exists, mention that briefly.
Always include file paths so the user can navigate to the source:
Don't overwhelm with every possible connection. Prioritize:
Present what you find, don't synthesize for the user.
Good:
"You've written about this in three places: [quotes]. What stands out to you?"
Avoid:
"Based on your notes, I think the pattern here is..."
If something might be relevant but you're not sure:
People appear in various contexts. Search for:
Pattern: "Roman|Dzeciol"
Pattern: "Daniel|Weiner"
Pattern: "Ryan M|arm injury" # Context-specific
Projects have codenames, formal names, and related concepts:
Pattern: "Epic Social|EGSM|Social POC"
Pattern: "EOS SDK|SDK team"
Pattern: "Epic-Social-39" # Directory name format
Daily notes follow naming conventions:
# This week (January 2026)
Glob: "**/00_Inbox/2026-01-2*.md"
# Last month
Glob: "**/00_Inbox/2025-12-*.md"
# Specific day
Glob: "**/00_Inbox/2026-01-15.md"
For relative timeframes, calculate the date range first.
Technical and abstract concepts:
Pattern: "DPoP|token exchange"
Pattern: "authentication|auth"
Pattern: "overlay|communication architecture"
User: "I need to think about what to do with Roman's situation."
Claude (research mode):
Grep pattern="Roman" path="/vault"Looking at your recent notes, Roman has been focused on the Epic Social Overlay deep-dive. From your January 20th note:
"Roman doing weekend deep-dive to front-load uncertainty; outcome Monday shapes next sprint."
And earlier on January 15th you noted he was handling the overlay communication architecture. What specifically are you thinking about regarding his situation?
User: "What happened this week with the POC?"
Claude (research mode):
Glob pattern="**/00_Inbox/2026-01-2*.md"From your notes this week:
Monday (Jan 20): You noted that Roman's weekend work would shape the sprint direction.
Tuesday (Jan 21): Discussion about auth being "good enough for POC"
(from
00_Inbox/2026-01-20.mdand00_Inbox/2026-01-21.md)What aspect are you trying to understand?
User: "Help me think through the authentication approach."
Claude (research mode):
Grep pattern="auth|DPoP|token" path="/vault"You've written about this in a few places. From your team status notes:
"Auth resolved toward 'good enough for POC' (token exchange short-term, DPoP long-term)"
And in the project README you have a section on auth requirements.
(from
02_Areas/EOS-SDK-Team/status.mdand01_Projects/Epic-Social-39/README.md)What's prompting the reconsideration?
User: "You don't need to keep searching my notes for this. Just answer directly."
Claude:
Understood. I'll respond conversationally without proactive vault searching. Let me know if you want me to look something up specifically.
Research mode complements other vault workflows:
The distinction: task management queries structured data (checkboxes). Research mode fetches unstructured context (prose, decisions, discussions).
Research mode transforms Claude into a proactive research assistant that:
The goal is to reduce the friction of finding relevant context while preserving the user's role in making sense of it.