// the find
danielmiessler/LifeOS
⛰️ LifeOS — The universal AI Harness designed to move you from Current to Ideal state in both life and work.
LifeOS is a huge bundle of markdown system prompts, TypeScript tooling, and shell scripts that turns Claude Code (or in theory any agentic harness) into a personal assistant with persistent memory, custom skills, and a self-modifying 'constitution'. It's aimed at power users who already live in an AI coding agent and want it to also manage their goals, contacts, and daily workflow, not at people looking for a conventional library or app.
Ships as one self-contained directory (a single 'skill') rather than scattering config across the filesystem, which makes install/uninstall/update reasonably clean and traceable. The harness-agnostic bet is a real architectural choice, not just a claim — it's built on hooks, skill files, and context docs instead of one vendor's proprietary API, so it doesn't hard-lock to Claude Code even though that's the only tested path. Release process has actual automated gates (sanitizer-residue checks, license-font checks, dead-doc-pointer checks) that run before publishing, which is more rigor around 'don't leak personal data or unlicensed assets' than most projects in this space bother with. Iteration is fast and changelogged in detail across major version bumps rather than silent drift.
The install flow is 'paste a URL into your agent and let it read + execute the setup' — that's an LLM agent fetching remote instructions and getting filesystem write access based on them, with 'asks permission before it touches anything' as the only stated safeguard; there's no diff-and-review step shown before changes land. The project layers on a large amount of bespoke terminology (Cortex, Synapse, Atlas, Ledger, Arbol, Bunker, Hermes, Pulse, ISA, the Algorithm) for concepts that mostly map to ordinary things — memory, event routing, asset inventory, changelog, sandboxing, a secondary CLI, a dashboard, a structured doc format — and that vocabulary is a real learning-curve tax for something that's ultimately prompt and config engineering. Since the public repo is generated from a private source tree, community PRs are described as 'ported... with credit rather than merged directly,' meaning outside contributors never actually land a commit that's theirs, and there's a single-maintainer bottleneck plus reduced auditability of true history. Most of the system's 'behavior' lives in tens of KB of natural-language doctrine steering an LLM, which by nature can't be unit tested the way the TypeScript tooling around it can — there's no way to verify the agent actually follows the constitution reliably beyond manual testing.