██████╗ ██████╗ ███████╗███████╗ █████╗ ██████╗  ██████╗██╗  ██╗
██╔════╝██╔═══██╗██╔════╝██╔════╝██╔══██╗██╔══██╗██╔════╝██║  ██║
██║     ██║   ██║███████╗█████╗  ███████║██████╔╝██║     ███████║
██║     ██║   ██║╚════██║██╔══╝  ██╔══██║██╔══██╗██║     ██╔══██║
╚██████╗╚██████╔╝███████║███████╗██║  ██║██║  ██║╚██████╗██║  ██║
 ╚═════╝ ╚═════╝ ╚══════╝╚══════╝╚═╝  ╚═╝╚═╝  ╚═╝ ╚═════╝╚═╝  ╚═╝

Co-research with AI. A ready-to-clone scaffold for research projects built around a simple idea: AI loses all context when a session ends. Cosearch solves this by turning Obsidian into a structured external memory store that both you and the AI read and write together.

Clone this repo, launch Claude Code, and the AI walks you through the rest.

ADVISOR.md (project root) defines how the AI behaves as a research advisor — it prioritizes research direction over execution details, and will tell you to scrap the whole approach if needed.

---

The problem this solves

AI is a powerful research partner — but it forgets everything between sessions. Without a shared memory, every session starts from scratch: re-explaining the project, re-establishing context, re-recovering where you left off.

Cosearch treats the Obsidian vault not just as a note-taking tool, but as the AI's long-term memory. Every session end writes a structured log. Every session start reads it back. Over time, the AI accumulates the context needed to work as a genuine collaborator.

---

What's included

Code / Data / Results (code/, data/, results/)

Research-oriented codebase structure:

Folder	Role	Git
code/scripts/	Run scripts	✅
code/pipelines/	Pipeline configs (*.yaml)	✅
code/notebooks/	Exploratory notebooks	✅
code/cosearch/	Cosearch framework package	✅
code/tests/	Test code	✅
code/legacy/	Pre-cosearch code archive	❌ local only
config/sessions/	Session config files	✅
config/params/	Analysis parameters	✅
data/examples/	Small sample data	✅
data/raw/	Raw input data	❌ local only
data/processed/	Preprocessed data	❌ local only
data/legacy/	Pre-cosearch data archive	❌ local only
results/final/	Final figures & tables	✅
results/runs/	Script outputs — runs/YYMM/YYMMDD_topic/	❌ local only
results/legacy/	Pre-cosearch results archive	❌ local only
results/logs/	Run logs	❌ local only
results/tmp/	Intermediate outputs	❌ local only

artifacts/manifest.yaml tracks all large local files outside Git — path, origin, and how to recreate them.

새 대용량 파일이 생길 때 할 일:

1. artifacts/manifest.yaml에 항목 추가 — local_path, note, created_by 기록
2. 관련 EXP 노트 Output 섹션에 경로 명시
3. results/final/에 넣을 최종본만 Git에 추가

Every analysis session links code → data → results through an EXP note in obsidian/50_Experiments/.

Obsidian vault (obsidian/)

Structured knowledge base readable by both human and AI:

Folder	Purpose
00_Meta/	Project status, conventions, correction log, understanding checklist
10_Literature/	Paper notes (citekey-based)
20_Concepts/	Theory and concept notes
30_Methodology/	Analysis pipeline and best practices
40_Projects/	Ongoing analysis notes
50_Experiments/	Session logs (EXP) and meeting notes (MTG)
60_Plans/	Roadmaps and next steps
70_Reviews/	Deep research results (DR) and code reviews (REV)
99_Assets/	Binary files only (images, PDFs, HTMLs)

Comes with Dataview pre-installed and graph view color-coded by folder.

Claude Code setup (.claude/)

Skills:

• session-wrap — End-of-session ritual: writes an EXP note, updates project status in Obsidian, and syncs the Claude memory snapshot
• organize-deepresearch — Absorbs a deep research .md file into the existing note structure rather than creating a separate summary doc
• advisor — Spawns an isolated agent to diagnose research direction against ADVISOR.md principles (layer 1→2→3) and saves the review to 08_Reviews/
• write-deepresearch-prompt — Writes a self-contained prompt doc to hand off to an external AI (Claude.ai deep research etc.); never executes the search itself
• teach — Turns a concept you don't understand into a structured note in 20_Concepts/; never explains in chat
• sync-from-cosearch — Fetches skill updates from this template repo and applies them selectively; warns at session start if more than 14 days since last sync
• obsidian-skills-main — Obsidian markdown, canvas, bases, and CLI helpers

session-wrap now includes code execution tracking: when you run an analysis, the EXP note records the script, input data, command, and output paths — linking every result back to the code that made it. New large local files get logged to artifacts/manifest.yaml.

Stop hook (check_memory_update.py): Fires automatically when Claude Code exits. Warns if today's EXP note is missing, the status file is stale, or the memory snapshot exceeds 50 lines.

settings.local.json: Pre-approved permissions for WebSearch, PubMed, and common academic publisher domains.

Memory system (memory/)

Starter MEMORY.md index for Claude's persistent project memory (stored at ~/.claude/projects/[encoded-path]/memory/).

---

Getting started

# 1. Clone
git clone https://github.com/dkfmaekdnjfk/Cosearch.git my-project
cd my-project

# 2. Fresh git history
rm -rf .git && git init

# 3. Open obsidian/ as a new vault in Obsidian

# 4. Launch Claude Code — it handles the rest
claude

On first launch, Claude reads obsidian/00_Meta/START_HERE.md and walks you through the remaining setup.

---

Daily workflow

When	What
Session start	"오늘 뭐하지?" — Claude reads status and picks up where you left off
Got deep research results	/organize-deepresearch — absorbs into vault
Need a research direction check	/advisor — isolated agent diagnoses direction against ADVISOR.md
Need to delegate a literature search	/write-deepresearch-prompt — writes the handoff doc, never searches itself
Session end	/session-wrap — logs the session, updates memory

---

Key conventions

• EXP notes: EXP YYMMDD topic.md in 50_Experiments/
• Literature notes: Author YYYY keyword.md in 10_Literature/
• Deep research files: drop in 70_Reviews/ as DR YYMMDD topic.md, then /organize-deepresearch

See obsidian/00_Meta/CONVENTIONS.md for the full format reference.

---

License

MIT — see LICENSE. obsidian-skills-main by @kepano, also MIT.