The folder structure I use so Claude can navigate my whole business and pull the exact note in one query. Copy it, fill it, keep it fed.
From @jlee.mov · pairs with CAROUSEL-005.
Everything is one pipeline. A note moves 00 → 07
as it goes from raw thought to shipped. Location tells you (and the AI) the status — no status field to maintain, no "misc" folder to guess through.
my-vault/
├── CLAUDE.md ← the map (see §2)
├── 00-Inbox/ raw drops, unsorted — clear it weekly
├── 01-Ideas/ captured + shaped, not started
├── 02-Drafts/ in progress
├── 03-Published/ done / shipped
├── 04-Sources/ raw material, assets, references
├── 05-Competitors/ research on other people
├── 06-Hooks/ reusable openers / patterns
├── 07-Analytics/ what actually worked
└── _meta/ ops docs — the owner files (see §2)
Rename the folders to your work (a consultant might use 03-Delivered
, 04-Client-Files
), but keep two rules:
Numbered by stage— the number is the lifecycle position.** Same folders every project**— boring and predictable beats clever. The AI learns the shape once.
Why it works: a note in 02-Drafts
is in progress; in 03-Published
, it's done. You never tell the AI a note's status — it reads it off the path. The structure carries the state so the notes don't have to.
CLAUDE.md
sits at the top and is a map, not a dump. It doesn't hold your content — it points to the one doc that owns each question. The AI reads the map first, then opens only the file it needs. Context stays lean; answers stay grounded.
Give every recurring question exactly one owner file in _meta/
:
| Question | Owner file |
|---|---|
| What's next? | _meta/TODO.md |
| Where did I leave off? | _meta/STATE.md |
| What's the strategy / who am I? | _meta/PROJECT-STATE.md |
The rule that keeps it clean: one owner per question. If two files disagree, the owner wins — fix the other on the spot. Two files never quietly contradict each other, so the AI never gets a split answer.
Minimal CLAUDE.md
to start:
This is a map, not the content. Read the owner file, not everything.
## Who
[one line: who you are, what you sell, who for]
## Where things live
- What's next → _meta/TODO.md
- Where we left off → _meta/STATE.md
- Strategy / positioning → _meta/PROJECT-STATE.md
- Folders are numbered by stage: 00-Inbox … 07-Analytics
## Standing rules
- [anything the AI should never get wrong: voice, hard no's, who to never email, etc.]
A small memory/
folder of one-fact files, indexed by a MEMORY.md
list. These are the things worth never re-explaining — your pricing rule, a client's quirk, a decision you already made.
memory/
├── MEMORY.md ← index: one line per fact, links to the file
├── pricing-rules.md ← one fact
├── client-acme-prefers-email.md ← one fact
└── voice-no-emoji-in-body.md ← one fact
One fact per file keeps them easy to add, correct, and delete. MEMORY.md
is the index the AI scans to decide what's relevant.
The structure gets the AI to the right folder. Semantic search gets it to the exact note — by what the note means, not the words you happened to use.
Easiest: pointClaude Code(or Claude Desktop with a folder connected) at the vault. It reads and navigates the files directly.** Add real semantic search**when the vault gets big: a local embeddings-based note search (e.g.qmd
, or an Obsidian semantic-search / smart-connections plugin) indexes the folder so a question like*"how did I price the last client?"*returns the one right note ranked at the top — instead of 40 keyword matches.
Ask a question → it pulls the exact passage → it answers from your real history. That's the difference between a note app and a second brain.
None of this works if you stop feeding it. The setup is an afternoon; keeping notes in their stage is the actual work. A dead vault is just a prettier version of forgetting. Ten minutes at the end of each day — move what shipped, clear the inbox — is what makes it compound.
Built this? Reply to the DM and tell me what folder names you landed on — I read every one.