Quartermaster

An AI operating system that runs locally, owns your files, remembers everything, and learns from every correction. Built on Claude Code and Obsidian. Re-explaining context every morning is a stupid use of time.

Get Started What's Different GitHub

Quartermaster turns a markdown vault into a persistent, self-improving AI assistant. Seven layers: vault (files) + skills (workflows) + hooks (automation) + daemon (always-on) + memory (persistence) + search (retrieval) + self-improvement (learning). Each works independently. Together they compound. Full architecture →

graph LR
    A["Capture<br/><small>phone, voice,<br/>transcripts</small>"] --> B["Process<br/><small>15 skills,<br/>auto-routing</small>"]
    B --> C["Store<br/><small>vault, search,<br/>three-mode retrieval</small>"]
    C --> D["Learn<br/><small>calibration,<br/>self-improvement</small>"]
    D --> E["Act<br/><small>daemon, alerts,<br/>follow-ups</small>"]
    E --> A

    style A fill:#161b22,stroke:#5eead4,color:#e6edf3
    style B fill:#161b22,stroke:#5eead4,color:#e6edf3
    style C fill:#161b22,stroke:#5eead4,color:#e6edf3
    style D fill:#161b22,stroke:#5eead4,color:#e6edf3
    style E fill:#161b22,stroke:#5eead4,color:#e6edf3

Five stages, one loop. Everything flows through a markdown vault. Full architecture →

Who this is for

Built for people who

• Use AI for real work - whether ChatGPT, Claude, or Claude Code - and want something that compounds instead of resetting every session
• Manage multiple domains with different stakeholders, contexts, and communication styles
• Process meetings and transcripts and need decisions, actions, and insights extracted automatically
• Want an AI that works between sessions - processing your inbox, flagging slipping tasks, sending you a morning plan before you open your laptop
• Build AI systems and want battle-tested architecture patterns for memory, search, self-improvement, and agent orchestration

Whether you're new to Claude Code or pushing its limits, pick your starting point →

What's genuinely new

Three things I haven't seen anywhere else

🔄

It rewrites its own instructions Corrections log to a calibration file. Patterns that appear 3+ times graduate to permanent rules. After 2 months: 130+ suggestions extracted, 7 graduated to permanent rules, each eliminating a recurring friction point. Every change is auditable - the calibration log shows exactly what changed and why. You don't configure it. You grow it. See real examples →

🔴

Five-lens red team on every important document /challenge runs 5 independent agents in parallel: audience fit, logic gaps, vault contradictions, your known blind spots, and a pre-mortem. Independent execution means no anchoring - the pre-mortem doesn't know what the audience-fit lens found. Verdict + top 3 fixes in one pass.

📝

It remembers why you decided what you decided Every structural change logged with reasoning. Contradiction detection catches when you reverse a previous decision: "This reverses the decision to remove the budget table. Intentional?" Rejected alternatives become permanent constraints unless consciously overturned.

The operational backbone

And five things done well

🌙

It never sleeps A daemon runs 24/7 on a second machine. Hourly heartbeat processes your inbox, flags slipping tasks, sends a morning plan to your phone before you wake up. The concept: agency between sessions. My implementation: a Telegram bot on a spare Mac. Yours could be a cron job on a VPS.

🎙️

Meetings become permanent, searchable knowledge Drop a transcript, walk away. The system auto-detects the theme, extracts decisions and actions, routes everything to the right folder. The concept: autonomous transcript-to-knowledge pipeline. My implementation: MacWhisper + iCloud. Yours could be Whisper + Dropbox + any transcription tool.

📋

Tasks managed like a chief of staff Leverage scoring (Impact x Effort) trumps due dates. Strategic weighting from live priorities. Waiting items age and trigger auto-drafted follow-ups after 7 days. The concept: AI-prioritised tasks that understand what actually matters this week, not just what's technically next.

📱

Captures from everywhere, processes centrally Phone, voice, transcripts, Telegram, share sheets, email forwarding. Everything converges on one inbox that gets processed automatically. The concept: many capture points, one processing pipeline. The tools are swappable - the architecture isn't.

🔗

Connected to your actual work tools Gmail search from the terminal. Markdown to rich HTML clipboard for pasting into Gmail, Word, or Outlook. Office document generation (DOCX, PPTX, XLSX). The output pipeline matters as much as the input - writing in markdown is pointless if you can't get it into the format stakeholders expect.

Read the full breakdown →

Get running in 3 steps

Quick start

# 1. Clone and copy the template
git clone https://github.com/sovrana/qm-os.git
cp -r qm-os/template/ ~/my-vault/ && cd ~/my-vault/

# 2. Customise CLAUDE.md (your name, your blind spots, your stakeholders)
# Search for CUSTOMISE - there are ~20 marked sections

# 3. Run your first morning plan
claude /morning

Needs: Claude Code + Python 3.10+ (for search) + Git. Full setup with semantic search and hooks takes 30 minutes →

A typical day

What a day looks like

6:30am - phone buzzes

Telegram: your morning plan is ready. 3 P1 items, 2 follow-ups auto-drafted for stale waiting items. A task on Alex is now at 12 days. You haven't opened your laptop.

9:00am - open Claude Code

The session-start hook loads a dashboard: tasks due today, items waiting on people, unprocessed inbox files, high-leverage quick wins. Yesterday's context is already in memory.

9:15am - /brief #project-a Alex

Last 3 meeting notes, open tasks, waiting items, stakeholder preferences. A "what NOT to say" section based on political context. Two minutes.

10:00am - meeting

MacWhisper recording the call. You focus on the conversation.

10:45am - drop transcript, walk away

The heartbeat auto-processes it: decisions, actions, insights extracted. Actions land in tasks.md tagged to the right theme. You didn't run a single command.

11:30am - /challenge the board paper

Five parallel lenses tear it apart simultaneously. Verdict: Needs Work. Top issue: execution mechanics missing (your known blind spot, caught automatically). Three fixes pushed to your task list.

2:00pm - /draft linkedin

Voice-calibrated against your real writing samples. Anti-slop enforced. 8 variants generated. Pick one. Copy to clipboard as rich HTML. Paste into LinkedIn.

5:00pm - session ends

Auto-commit to git. Search reindexes. Nothing lost. Telegram: evening summary. 5/7 planned items done. Tomorrow's top 3. That waiting item on Alex is flagged for escalation.

Sunday - /weekly

7 parallel subagents: task audit, stale item cleanup, memory refresh, cross-theme connection discovery, self-improvement suggestions, stakeholder heatmap, decision digest. The system gets smarter every week.

/challenge tearing apart a strategy document - 5 parallel lenses, verdict in 2 minutes

Type a command, get a result

15 skills included

Command	What it does	Time
/morning	Prioritised daily plan with leverage scoring	~5 min
/brief	Pre-meeting one-pager with political context	~2 min
/challenge	5-lens parallel red team on any document	~5 min
/stress-test	3-persona adversarial debate	~15 min
/transform	Transcript to structured knowledge	~5 min
/draft	Voice-calibrated outbound content	~5 min
/prep	Vault research before writing	~10 min
/prompt	Structure messy thinking into executable briefs	~30 sec
/capture	Stress-test and store rhetorical weapons	~30 sec
/inbox	Process captures and route to tasks	~5 min
/show	Filtered task views by theme or person	~30 sec
/weekly	Full system audit with 7 parallel subagents	~30 min
/changelog	Iteration decisions with contradiction detection	~1 min
/visualise	Infographics and diagrams from vault notes	~2 min
/evening	End-of-day reflection and pattern tracking	~5 min

I've built more for my own workflows: /gmail, /publish, /clipboard. Writing a new skill takes under 10 minutes. How skills work →

Browse all skills →

Quickstart (30 min) → Architecture deep dive → View on GitHub

Built with Claude Code and Material for MkDocs. Something broken? Open an issue or start a discussion.