A Lightweight Operating System for Long-Horizon Human–AI Collaboration
If you've worked with an AI assistant for more than a few days, you've probably noticed something:
- It forgets what you agreed on last week
- Numbers drift or get recalculated differently each time
- You end up re-explaining context every session
- Small misunderstandings compound into confusion
- What felt like a partnership starts feeling unreliable
This isn't a prompting problem. It's a structural problem.
AI systems don't retain memory across sessions by default. They optimise for fluent responses, not consistent ones. Without external structure, long-term collaboration quietly degrades — not through dramatic failure, but through slow drift.
LC-OS (Lean Collaboration Operating System) is a minimal set of practices and templates that stabilise long-horizon human–AI work.
It's not software. It's not a platform. It's a way of organising your collaboration so that:
- Memory persists — through documents you maintain, not model memory
- Truth has a home — one authoritative source for decisions and numbers
- Failures become visible — logged and traceable, not hidden
- Repair is structured — clear steps to recover when things break
The core insight: reliability comes from governance, not capability. A well-structured collaboration with a standard model outperforms an unstructured one with a frontier model.
LC-OS is for practitioners who:
- Work with AI assistants on projects spanning weeks or months
- Handle domains where accuracy matters (finance, research, planning, strategy)
- Have experienced drift, confusion, or trust breakdown in long collaborations
- Want a lightweight system, not heavy infrastructure
It's not for:
- Single-session tasks or quick queries
- Fully automated pipelines with no human involvement
- Those seeking a plug-and-play software solution
Three templates. One page of guidance. Get running in 30 minutes.
Best if you want to test the approach before committing.
Complete toolkit with worked examples, repair protocols, failure logging, and governance rules.
Best if you're serious about sustained collaboration and want the full system.
Running Document
A persistent file that captures decisions, rules, and corrections. Read by you and the AI at the start of each session. This is your shared memory.
Canonical Numbers
One source of truth for all numerical data. The AI references it; it doesn't recalculate from memory. Eliminates a whole class of errors.
Failure Logging
When something breaks, you log it: what happened, why, how it was fixed. Failures become learning, not embarrassment.
Repair Protocol
A simple sequence when things go wrong: Stop → Diagnose → Rollback → Note. No drama. Just structured recovery.
Stability Ping
A brief check-in after major milestones: Are we still aligned? Any drift? One improvement before continuing?
- Choose your path — Minimal or Full
- Copy the templates — download or fork this repo
- Start a session — share the Running Document with your AI at the start
- Work normally — but log decisions and update the document as you go
- When things break — use the repair protocol instead of pushing through
That's it. The system is lightweight by design.
- It won't make your AI smarter
- It won't prevent all errors
- It won't work if you don't maintain the documents
- It's not magic
What it does do: create conditions where errors are visible, contained, and repairable — so that long-horizon collaboration can actually sustain itself.
This toolkit is based on a series of research papers documenting a year-long human–AI collaboration. Reading them is not required to use LC-OS, but if you want the theory behind the practice:
CC BY 4.0 — Use freely, adapt as needed, attribution appreciated.
This is a living project. If you adopt LC-OS and develop improvements, variations, or domain-specific templates, contributions are welcome.
Stability is not the absence of failure; it is the capacity for visible, structured repair.