company-ai-os
CLAUDE.md2 starsChanged 3 months agoMITView on GitHub
This vault is the company's second brain and AI Operating System. It is a hub that maps every department, project, and tool the company runs on, and can route a task into the right real folder or system and act there. Any agent (you) that opens this folder MUST read this file first, then follow it exactly. Built on Karpathy's LLM-wiki idea (raw → wiki → schema; ingest/query/lint) plus a dispatch layer that turns the wiki into an OS. Reference:…
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- eodSkill
- lintSkill
- new-entitySkill
- onboardSkill
- session-handoffSkill
- session-startSkill
The file
Show its rules and text
We split this CLAUDE.md into the rules it gives an agent
21 rules. About 0 are common, 9 are distinctive, and 12 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈9
Distinctive · estimate
This author's own choices. Worth a look.
12
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 47
Use aliases: in frontmatter for every other name the same thing goes by.
- Very specific: line 53
Every note in entities/ uses this frontmatter + body:
- Distinctive: line 85
Every interaction is one of these.
- Very specific: line 88
Read the source (a raw/ file, a URL, or a real folder/system export).
- Very specific: line 90
Write/update the relevant entities/ note(s) with the extracted facts.
- Very specific: line 91
Add or update the entry in index.md.
- Distinctive: line 97
Read them, synthesize an answer, cite the note(s) you used.
- Distinctive: line 98
If the answer is valuable and reusable, file it back as a new/updated note (mini-ingest) and log it.
- Distinctive: line 110
If ambiguous or unmapped, ask which entity — do not guess a folder.
- Very specific: line 112
Locate: Read that note's path:. That is the working directory for the task.
- Very specific: line 113
Act: Read freely inside it. Before creating/editing/deleting any file there, confirm with the user (see §5).
- Very specific: line 115
Record: After acting, update the entity note's current-state + updated:, and log: ## [date] dispatch | <task> in [[Entity]] → <what …
- Distinctive: line 126
Never guess a path: If no entity resolves, ask.
- Distinctive: line 127
Never assume a price: When quoting any tool, subscription, plan, or vendor in a proposal or budget, fetch the current price (web) first and …
- Distinctive: line 129
No stale or guessed figures.
- Very specific: line 133
Log every operation: If it's not in log.md, it didn't happen.
- Very specific: line 134
Sources are immutable: Never edit files in raw/ after ingest.
- Distinctive: line 141
When a new session opens this vault:
- Distinctive: line 142
Read this file (CLAUDE.md).
- Very specific: line 143
Read mistakes.md — standing rules from past failures.
- Very specific: line 145
If .claude/sessions/LATEST.md exists, read it — that is the handoff from the previous session (or run /session-start).
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Company AI OS — Operating Manual (Schema)
This vault is the company's second brain and AI Operating System. It is a hub that maps every department, project, and tool the company runs on, and can route a task into the right real folder or system and act there. Any agent (you) that opens this folder MUST read this file first, then follow it exactly.
Built on Karpathy's LLM-wiki idea (raw → wiki → schema; ingest/query/lint) plus a dispatch layer that turns the wiki into an OS. Reference: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
1. The three layers
- Raw sources — immutable inputs. Live in
raw/(andraw/assets/for attachments), OR they are the real working folders/systems outside this vault (shared drives, department folders, SaaS tools). - The wiki — LLM-maintained markdown.
index.md,log.md, and one note per entity inentities/. The wiki is the map. The[[wikilinks]]between notes ARE the graph. - The schema — this file. Rules for how the wiki is built, queried, linted, and dispatched.
2. Folder conventions
company-ai-os/ ← this vault (the hub)
├── CLAUDE.md ← this schema (read first, always)
├── index.md ← catalog of every wiki page, by category
├── log.md ← append-only history of every operation
├── entities/ ← one note per entity = the graph nodes
│ ├── Company.md ← the company itself ("who are we")
│ ├── HR.md · Engineering.md · Operations.md · Marketing.md · Finance.md · Sales.md
│ ├── Routing.md ← name → folder/system routing table
│ └── <Team/Project/Tool>.md ← grows over time
├── raw/ ← dropzone: paste docs, notes, exports to ingest
│ └── assets/ ← images, PDFs, binaries referenced by raw notes
├── _templates/ ← note skeletons (entity, ingest)
└── .claude/ ← skills (slash commands) + department sub-agentsThe real work may live OUTSIDE this vault — department folders, shared drives,
repos, SaaS tools. Entity notes carry the exact path: (or system link) to that real location.
Naming: entity note filename = the entity's common name (spaces OK).
Use aliases: in frontmatter for every other name the same thing goes by.
3. Entity note format
Every note in entities/ uses this frontmatter + body:
---
type: company | department | team | project | tool | person | reference
path: "../Engineering" # real folder, forward slashes; "" if pure concept or SaaS
aliases: [every, other, name, this, goes, by]
tags: [department, active]
status: active | paused | archived
updated: 2026-01-01
---
# <Name>
**What it is** — one line.
**Where** — `path` above, or links to the systems it lives in. Key subfolders that matter.
## Details
...facts, decisions, current state...
## Links
[[Company]] · [[related entity]] ← these build the graphRule: link liberally. A [[Name]] that has no note yet is fine — it marks a node
worth creating later.
4. The four operations
Every interaction is one of these. Name the operation you're doing in your reply.
INGEST — add a source to the brain
- Read the source (a
raw/file, a URL, or a real folder/system export). - Discuss the key takeaways with the user (1–3 lines) — don't silently absorb.
- Write/update the relevant
entities/note(s) with the extracted facts. - Add or update the entry in
index.md. - Append one line to
log.md:## [YYYY-MM-DD] ingest | <source> → <notes touched>. - Cross-link with
[[wikilinks]].
QUERY — answer from the brain
- Find relevant
entities/notes (start atindex.md). - Read them, synthesize an answer, cite the note(s) you used.
- If the answer is valuable and reusable, file it back as a new/updated note (mini-ingest) and log it.
LINT — health-check the brain
Scan for: contradictions between notes · stale claims (updated: old + source changed) ·
orphan notes (no [[links]] in or out) · dead path: (folder moved/gone) ·
gaps (entity mentioned but no note).
Report findings; fix with confirmation; log as ## [date] lint | N issues → M fixed.
DISPATCH — the OS layer: do the thing in the right place
- Resolve. Map the request to an entity: match name/alias against
entities/Routing.md(the name→folder table), thenentities/andindex.md. If ambiguous or unmapped, ask which entity — do not guess a folder. (Orchestrating an agent team? SeeAGENTS.md.) - Locate. Read that note's
path:. That is the working directory for the task. - Act. Read freely inside it. Before creating/editing/deleting any file there, confirm with the user (see §5).
- Record. After acting, update the entity note's current-state +
updated:, and log:## [date] dispatch | <task> in [[Entity]] → <what changed>.
5. Guardrails (non-negotiable)
- Read freely, confirm writes inside real folders outside this vault. Reading/mapping any real folder = always allowed. Creating/editing/deleting a file inside a department or project folder = state the exact path + change, then wait for OK. Writes inside the vault itself do not need confirmation.
- Never guess a path. If no entity resolves, ask. A wrong folder is worse than a question.
- Never assume a price. When quoting any tool, subscription, plan, or vendor in a proposal or budget, fetch the current price (web) first and note the date checked. No stale or guessed figures.
- Never expose sensitive data. Payroll figures, personal employee data, credentials, and customer PII stay in their source systems — reference them in notes, never copy the values into the vault.
- Log every operation. If it's not in
log.md, it didn't happen. - Sources are immutable. Never edit files in
raw/after ingest; supersede with a new note instead.
6. Session start behavior
When a new session opens this vault:
- Read this file (
CLAUDE.md). - Read
mistakes.md— standing rules from past failures. They bind you. - Skim
index.mdfor the current map andlog.md(last few entries) for recent state. - If
.claude/sessions/LATEST.mdexists, read it — that is the handoff from the previous session (or run/session-start). - First session ever (vault still contains Acme sample data)? Offer
/onboard. - Then act on the request as one of the four operations above.
7. The learning loop
Two append-only files at the vault root keep the OS from repeating itself:
learnings.md— what worked; reusable wins with a "rule going forward".mistakes.md— what failed; root cause + a rule so it never happens twice.
Rules:
mistakes.mdis read at EVERY session start (§6). Its rules are binding until the user retires them.- Whenever something breaks or the user corrects you, append to
mistakes.mdin the same session — don't wait for/eod. /eodand/session-handoffboth ask: "anything learned today?" and file it.- A mistake that keeps recurring graduates into this schema (§5) — propose the edit.
8. Writing on the company's behalf
Any outward-facing text (email, WhatsApp, LinkedIn, proposals) follows
entities/Style Guide.md — per-channel voice rules with worked samples. Load it
before drafting; the creative-director agent reviews against it. Sending anything
externally always requires explicit user confirmation.
Schema v1 · Company AI OS template · grow this file as your OS matures.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

