The useful part of a personal LLM wiki is not that it stores a lot of text.
Storage is the easy part. Chat transcripts, notes, exported documents, project plans, and issue threads accumulate quickly. The harder problem is deciding which parts of that history should influence future agent behavior, which parts are stale, and which parts should never leave a private boundary.
That is the operating model I have been building toward: raw notes become source notes, source notes become curated wiki pages, curated pages become context packs, and context packs become agent-readable memory.
Reliability engineering often starts with tooling, but the durable work starts earlier: naming what actually fails.
That sounds smaller than it is. Names are the unit that lets an organization learn from production. Without them, incidents remain stories: this deploy, that overload, this dependency outage, that missing alert. The stories may be accurate, but they are hard to compare. They depend on who was in the room, who wrote the postmortem, and which details were memorable under stress.
I have written before about reef-pi as a public engineering artifact: modules, documentation, community, hardware. That post surveyed the shape of the project. This one is about a single line inside it.
The difference between a project and a platform is not size, and it is not popularity. It is whether there is a boundary somewhere in the system that the maintainers agree not to cross, and whether that agreement survives contact with real hardware.
The first time I cared deeply about infrastructure as code, the argument was not really about Chef.
Chef was the tool in front of us. It gave us cookbooks, recipes, resources, and a DSL for describing how machines should be configured. But the deeper argument was that operational knowledge should not live only in tickets, shell history, wiki pages, and the heads of a few senior engineers.
If a server needed a package, file, daemon, permission, or dependency, that intent belonged in code. If a change could break production, it deserved version control, review, testing, and a release path. If an operator repeated the same judgment, the system should eventually learn it as an artifact.
The more I use coding agents and LLMs for real work, the less I think of prompts as throwaway text.
A useful prompt is not just a question. It is a small interface. It carries assumptions, constraints, examples, preferences, and failure modes. A project instruction file is even more explicit: it tells an agent how to navigate a codebase, what quality bar to hold, which tools to prefer, and which boundaries not to cross. A context pack does the same thing for a domain. It packages durable facts, active goals, source notes, open questions, stale claims, and sensitive exclusions so an agent can reason from the right starting point.