PEPPERCORN

Posted on Jul 25

[Day 18] I set up a company of AI agents with Claude Code — and a local LLM (qwen) joined as the caretaker

#localllm #ai #dgxspark #agents

Intro

Day 18!

In Day 17 I audited 33,469 of my own AI conversations. Building on that, today I set up a company of cats to run my task management 🐱

What I used: Claude Code / DGX Spark / Slack & Notion.

Today's plan

What I want: hand the annoying parts of task management to a company of cats
Approach: build in order — the org, the task flow, the ledger, AI collaboration rules, then deploy to both machines
Done means: every morning, Necco and I pick "today's three tasks"
Result:
- A company of six cats. Inbox = Slack, source of truth = Notion, memory = the ledger
- I inventoried my projects: 80 of them. All in the ledger now
- One page of collaboration rules for Claude Code / Codex / qwen

① Built the org

Role-based staff (subagents), all named "___-neko" (neko = cat).

Name	Role
Necco	Chief secretary: morning meeting, task triage, routing, ledger upkeep
Shirabe-neko	Research
Tsukuri-neko	Implementation
Mekiki-neko	Review (read-only)
Soroban-neko	Number crunching
Rusuban-neko	Caretaker: a local LLM (qwen) living on the DGX. Summarizes, sorts, and tags incoming files

Before building, I looked at prior art. An individual delegating their own management to AI seems established as the AI Chief of Staff pattern.

The staff internals come tomorrow in Day 19.

② Built the task flow

A task pops up → one line into Slack from my phone (3 seconds)
             → [morning meeting] Necco sweeps the inbox
             → cleans it up, adds a due date, registers it in Notion
             → proposes "today's three"

Think of something, drop one line in Slack. By morning it's in Notion. An evening "done: ___" one-liner feeds the next morning's ledger update and task close-out.

③ Built the ledger

What's the ledger (LEDGER.md)? One file listing every project, so Necco can answer "where was that project again?" instantly.

For the initial data, I inventoried both of my machines. The count: 80 projects. I had no idea it was that many...

④ Set the AI collaboration rules

Instead of deciding "which AI gets this job" every time, it's now one page:

AI	Role
Claude Code	Lead (design, planning, review, conversation)
Codex CLI	Routine subcontractor (implementation)
qwen (local LLM)	Batch pre-processing (summarize, sort, tag — stays on the DGX)

Rusuban-neko is the DGX side of this rule. Zero API cost, and the data never leaves the house.

⑤ Deployed to both machines

The company is a git repository. I gave it a private GitHub remote and cloned it onto the DGX. An install script places the staff and skills into ~/.claude/.

Now either machine can summon the same company.

The first morning meeting

I opened Claude Code in the office and ran the first morning meeting.

The one-liner I had tossed into Slack the night before was cleaned up and registered in Notion by Necco. This is genuinely comfortable.

The full picture


Chief secretary	Necco (Claude Code)
Staff	Shirabe / Tsukuri / Mekiki / Soroban (subagents)
Caretaker	Rusuban-neko (local qwen on the DGX)
Inbox	Slack
Source of truth	Notion
Memory	the ledger file (80 projects)

Layer	Role
Slack	The inbox — ephemeral, free-form, 3 seconds to post
Notion	The source of truth — same as before
Ledger	The company memory — what exists, where, in what state

secretary/                   ← the office (cloned on both machines)
  CLAUDE.md                  ← Necco's persona, house rules, meeting runbook
  LEDGER.md                  ← the ledger (80 projects)
  input/                     ← file inbox (pre-processed by Rusuban-neko)
  agents/                    ← staff definitions (source of truth)
  skills/                    ← skills (source of truth)
  install.sh / install.ps1   ← places staff & skills into ~/.claude/

Notion stays; only the entrance changed. Necco does the formatting and registering.

Today's takeaways

Separate the inbox from the source of truth: Notion felt tedious because it was doing both jobs
Keep the company's memory in git: AI auto-memory splits across two machines
Default patterns are fine: borrow the shape from prior art, fit the details to your own setup

The details

The morning meeting runbook (8 steps)

git pull (bring in the other machine's updates)
Rusuban-neko pre-processes anything new in input/
Close out yesterday's three (update Notion statuses)
Sweep the Slack inbox → turn into TODOs → register in Notion
Notion + calendar + inbox + ledger → propose "today's three"
Reflect the evening one-liners into the ledger
Post the final digest to Slack
git push

Git sync (1 and 8) and close-out (3) are baked into the runbook. The human only "posts" and "answers".

Pitfalls I mapped out beforehand

Pitfall	Countermeasure
The ledger goes stale	The evening one-liner habit
AI memory splits across two machines	Permanent knowledge goes into git-tracked files; auto-memory is treated as cache
Forgetting to sync the two machines	pull / push are steps 1 and 8 of the runbook
Work calendar isn't visible	Fill in verbally at the morning meeting, for now
No closed loop on task completion	The meeting starts with closing out yesterday's three

Two-machine sync

Staff definitions (agents/) and skills (skills/) live in the repository as the source of truth
install.sh (Linux) / install.ps1 (Windows) place them into ~/.claude/
Adding a staff member = commit → pull on the other machine → run install

Prior-art notes

mimurchison/claude-chief-of-staff: CLAUDE.md at the core; goals filter every priority; humans keep the final say
Claudia: remembers promises and relationships; a messaging gateway from your phone
loganhc-09/claude-chief-of-staff: resident operation via scheduled scripts

The common rule: always-on rules → CLAUDE.md / occasional methods → skills / context-heavy work → subagents. This build follows it too.

Tomorrow: Day 19

Staff day — the four subagent definition files, plus getting Rusuban-neko (the local LLM) running as a resident service.

Thanks for reading!

DEV Community