Skip to main content

Minimum Viable Jarvis

The open-source starter template for your Personal Agentic OS. One click to get your own private copy. Open it, run one command, and your Jarvis walks you through the rest.

What it is​

The Minimum Viable Jarvis (MVJ) template is a structured workspace of plain markdown files that gives you AI-augmented recall, strategic clarity, and compounding context, from the first session forward. Your harness (Claude Code, Hermes, Codex, or whichever wins this quarter's tradeoff) reads the files in this workspace and operates from them. Every conversation makes the system smarter.

This is the on-ramp. You do not start from a blank folder. You start from a structure that already knows where people files go, where artifacts go, where transcripts go, where your skills live. By the time you finish your first session, you have a profile, a plan for your most important blocker, and a workspace your AI can keep building on.

The repo is public, free, and forkable: github.com/Applied-AI-Society/minimum-viable-jarvis.

What lives inside​

your-jarvis/
├── CLAUDE.md                    # Points to AGENTS.md
├── AGENTS.md                    # Operating instructions for any AI agent
├── README.md                    # Setup + structure overview
├── user/
│   └── USER.md                  # Your profile (created on first session)
├── people/                      # One file per person in your life
├── artifacts/                   # Strategic documents, plans, decisions
├── meeting-transcripts/         # Processed transcripts from meetings
├── scripts/                     # Workspace automation
└── .agents/
    └── skills/                  # Canonical skills location

AGENTS.md is the instruction file the harness reads first. It tells your agent how to route inputs (a brain dump goes here, a transcript goes there, a person mention updates that file). CLAUDE.md is a thin pointer to AGENTS.md so Claude Code picks it up automatically.

The four data folders (user/, people/, artifacts/, meeting-transcripts/) are the substrate your agent reads from and writes to. Plain markdown. On your own disk. Portable across harnesses. [See: Skill File First, App Second for why this is the architecture.]

What ships in the skills folder​

The template ships with a working library of skill files under .agents/skills/. Each one is a markdown file the harness can run as a slash command (/onboard, /sync-with-upstream, etc.) or as a natural-language phrase ("onboard me", "sync with upstream"):

SkillWhat it does
/onboardFirst-session flow. Imports your AI history, builds your profile, interviews you on your most important blocker.
/create-user-profileBuilds or updates your profile through interview.
/get-unlockedStrategic interview when you are stuck. Leads with the Unlock Question.
/audit-my-weekPulls last 7 days of activity, reflects against priorities, saves a retrospective.
/process-braindumpRoutes unstructured input (paste, voice transcript, meeting debrief) to the right files.
/decongestExtracts a single stuck thought from your head into a clean artifact. The upstream practice that feeds every other skill.
/prep-for-meetingGenerates a meeting brief from people files, prior transcripts, and active projects.
/process-transcriptExtracts insights from a meeting transcript and updates the relevant people files.
/sync-granolaImports recent meetings from Granola into meeting-transcripts/.
/create-skillWalks you through creating a new skill of your own.
/sync-with-upstreamBacks up your work to your private GitHub repo and pulls upstream improvements from the MVJ template.

The skills folder lives at .agents/skills/ (vendor-neutral, the path Codex auto-discovers and Hermes recommends). For Claude Code, which auto-discovers .claude/skills/, the template ships a symlink so the same skills get picked up. You learn one skills folder. Every harness reads it.

How to start​

  1. Install the prerequisites: Node.js, Claude Code, VS Code, and the GitHub CLI.
  2. Sign in to GitHub from the CLI: gh auth login.
  3. Visit the MVJ repo and click "Use this template" → "Create a new repository". Name it, set it private, click Create.
  4. Clone your private copy locally. The convention is ~/Documents/github-repos/<repo-name>.
  5. Open it in VS Code: code <repo-name>.
  6. Open the integrated terminal (Cmd+`) and run clauded (or claude if you have not set up the alias yet, see the Post-Workshop Cheat Sheet).
  7. The first time you run it, your Jarvis automatically fires the /onboard skill. Follow the prompts. Drop your AI history (ChatGPT export, Claude export, LinkedIn) into user/ if you have any. Answer the interview questions.
  8. Turn on hourly auto-sync so your work is backed up to GitHub: bash scripts/install-sync-cron.sh.

By the end of step 7, you have a user/USER.md profile and an artifacts/<plan>.md for your most important blocker. The system knows who you are. Every future session starts there.

Why the template matters​

Three things change when you operate from a structured template instead of a blank folder:

1. The agent already knows the routing rules. Drop a brain dump and your agent already knows: people mentions update people/, strategic thinking goes to artifacts/, meeting recaps go to meeting-transcripts/, self-updates go to user/USER.md. You do not have to teach it where things live. The template did that already.

2. The skills you need on day one are already there. First session you can /onboard. End of week you can /audit-my-week. Stuck on something you can /get-unlocked. New person walks into your life you can ask the agent to update people/firstname-lastname.md and the convention is already set.

3. The template gets better over time and you can pull updates. As skills improve, as scripts get sharper, as AGENTS.md tightens, you pull those upstream improvements into your private workspace via /sync-with-upstream without losing any of your personal files. Your private workspace stays yours; the upstream template stays a moving baseline you can opt into.

Vendor-neutral by design​

The template does not couple you to one harness. The skills folder lives at .agents/skills/ because that path:

For Claude Code, which hardcodes auto-discovery to .claude/skills/, the template ships a symlink .claude/skills → ../.agents/skills so the same skill files appear automatically. [See: Learn The Harness, Not The Wrapper for why this matters.]

You can swap Claude Code for Codex or Hermes on day 30 and your context garden, your skills, your routing rules all come with you. Vendor lock-in at the harness layer is a 2024 pattern. The template was designed past it.

What the template is not​

  • Not a chatbot. The template is a persistent context substrate that compounds. A chatbot starts every conversation from zero. Your MVJ workspace starts every conversation from everything you have ever told it.
  • Not a SaaS product. No login, no subscription, no platform. It is markdown files on your disk that an AI reads.
  • Not opinionated about what you build. The template is shape, not content. What you put inside (your projects, your people, your plans, your skills) is yours.

Where to go next​