LLM Wikiで自分の「トリセツ」を作ってみたThe author explores LLM Wiki—a knowledge-design framework distinct from RAG and…

A recent approach to organizing information for large language models, known as LLM Wiki, is drawing attention as an alternative to more familiar retrieval methods, and one developer has now put it to a personal test by building an "owner's manual" — a トリセツ, or set of instructions — that captures their own values, tastes, and quirks. The experiment, documented on the Zenn platform, offers an early look at how individuals might structure knowledge about themselves so that a model can reason about it more reliably.

At its simplest, LLM Wiki is described as one way of designing the knowledge that a language model uses as the basis for its answers. In that respect it sits on the same continuum as Retrieval-Augmented Generation (RAG) and Graph RAG, both of which are widely used to give models access to information beyond their training data. Where LLM Wiki appears to differ is in its structure: rather than chunking documents into embeddings for similarity search, or mapping entities and relationships into a graph, it organizes knowledge in a wiki-like form that the author frames as distinct from those established techniques and from conventional structured data.

To understand why that distinction matters, it helps to recall how the adjacent tools work. RAG typically splits source material into passages, converts them into vector embeddings, and retrieves the most relevant fragments at query time to append to a prompt. Graph RAG extends this by building a knowledge graph, allowing a model to follow explicit connections between concepts rather than relying on semantic proximity alone. Both approaches improve grounding and reduce hallucination, but they can struggle with coherence when information is scattered across many small chunks or when relationships are left implicit. LLM Wiki is presented as a response to those limitations, favoring human-readable, interlinked pages over fragmented retrieval.

The framework is credited to a former OpenAI figure, though the concept still appears to be emerging rather than standardized, and readers should treat it as an evolving idea rather than a fixed specification. That caveat is worth keeping in mind, because much of the current interest comes from practitioners experimenting publicly, as the Zenn post itself demonstrates.

The author's chosen use case — a personal owner's manual — is a telling one. Instead of documenting a product or a codebase, they compiled their own preferences, values, and likes and dislikes into a structured reference the model can draw on. The appeal is that a well-organized profile of this kind could let an assistant respond in a more personalized and consistent way, without the user having to restate context in every conversation. It is a compact illustration of a broader trend often called personal knowledge management, in which people curate their own information so that tools can act on it.

This kind of self-directed structuring also connects to the wider shift from prompt engineering toward what some now call context engineering — the practice of deciding not just how to phrase a request, but what supporting information to place in front of the model and how to arrange it. Personal owner's manuals, memory features in consumer chatbots, and custom instruction fields are all attempts to solve the same underlying problem: models tend to perform better when the relevant context is deliberately shaped rather than left implicit.

For readers considering a similar project, the takeaway from the experiment is practical rather than definitive. LLM Wiki is likely most useful as a way to think about knowledge design, and its advantages over RAG or Graph RAG will depend heavily on the volume of information, the need for explicit relationships, and how the underlying model consumes the content. As with many nascent techniques in the local and applied LLM space, the clearest value may come from trying it on a small, well-scoped dataset — and a single person's tastes and values, as this account shows, make for a convenient starting point.