korab eland
← all tinkering
project in flight

Personal OS: learning by building on open source

I learn by doing. I used an open-source personal system to learn the basics of giving an AI agent lasting context, then built on top to make it my own: dual memory, custom automation, and cross-model checks.

role
Solo design + build
team
Solo project
stack
Claude Code · MCP · Python · git
span
2025
status
in flight

I learn by doing. Reading the docs or planning a system on paper never quite sticks for me. I understand something once I have built it, broken it, and built it again. So when I wanted a personal system for working with AI, I did not design one from scratch. I picked a working open-source one and learned on it.

That was PersonalOS by Aman Khan, the first repo I ever pointed Claude Code at and told it to set up for me. At its core it is plain: a few folders of markdown for goals, tasks, and a knowledge base, with a script that installs like a wizard. Using it taught me the thing I was actually after, how to give an AI agent context that lasts so I never have to repeat myself. It stands in for a project manager, captures context from email, meetings, and documents as it goes, gives me a morning brief of priorities aligned to goals, and every first draft starts inside it. None of that is unique. Plenty of second-brain setups look similar. What made it the right teacher was how little it asked to get started.

Once the basics were mine, I built on top, and each addition taught me the next thing. I added a dual memory, a short session file for immediate context and a longer-lived knowledge graph for durable decisions. Building the write-gate is how I learned why gating matters: nothing is stored without a freshness or similarity check, so the system absorbs new sessions without turning stale or duplicated. I wrote roughly a dozen skills for the jobs the base did not cover: a weekly brief that fans a subagent over each project, a cleanup pass that hands my session to a different model to catch its own confirmation bias, voice calibration that tunes drafts to how I actually write. I understood multi-agent review because I wired one, not because I read about it.

Months on, I have not needed a major change, which is the sign a simple system is working. The system and my understanding grew together. That is how I learn best: start from something real, then build on it until it is mine.

outcome

Original automation built on the base ~12 skills — how this was measured

Purpose-built Claude Code skills authored on top of the open-source foundation, beyond what the template shipped: a weekly alignment brief, cross-model review, memory sync, voice calibration, meeting and document ingestion, among others.

How the system holds context across sessions 2 horizons — how this was measured

A short, auto-loaded session file for immediate context plus a dedup-gated long-term knowledge graph for durable decisions. Nothing is written to either without a freshness or similarity check.

reflection

I do not learn much from documentation. I learn by building the thing, breaking it, and building it again. The open-source base gave me something real to start from, and everything I understand about context for AI agents I learned by extending it.