Skip to content

Template Philosophy

Jump to bottom
Chris Sweet edited this page Jul 14, 2026 · 2 revisions

type: synthesis up: "What-is-the-llm-wiki-Template" extends: "What-is-the-llm-wiki-Template" related:

Template Philosophy

The ideas behind the template's design, for readers who want to know why it works the way it does. Most of this page is grounded in the template's own documentation (llm-wiki.md, the template wiki's Wiki-LLM-Pattern and Three-Operations pages); the retrieval and competence-acquisition framing draws on the vision wiki, the reasoning trail behind the architecture, and is flagged as such where it goes beyond shipped template mechanics.

Why should the LLM write the wiki instead of me?

Because the expensive, value-destroying part of a knowledge base is the bookkeeping, not the thinking. Updating cross-references, keeping summaries current, noting when new data contradicts old claims, maintaining consistency across dozens of pages: humans abandon wikis precisely because this maintenance grows faster than the value. LLMs do not get bored, do not forget to update a cross-reference, and can touch 15 files in one pass. So the division of labor is: the human curates sources, directs analysis, asks good questions, and decides what it all means; the LLM does everything else. The wiki stays maintained because maintenance is nearly free.

What is the compounding-artifact argument for wikis over raw RAG?

A RAG pipeline over raw sources rediscovers, re-synthesizes, and re-explains the same things every session. There is no place for an insight to settle. The wiki is that place. Once a finding is filed, it is referenced by name rather than rediscovered, and cross-references compound: every new page that links to a finding makes it easier to surface next time. The wiki is also human-readable, version-controlled, and queryable, not a black box embedding index. The contribution is a place where insights live, a layer between the raw material and the conventions that describe how to work with it. See What-is-the-llm-wiki-Template for the same point aimed at newcomers.

What are the three operations (ingest, query, lint) and why those three?

The agent's interaction with the wiki is constrained to three named operations:

  • Ingest (write): after completing significant work, create or update pages, fix cross-references in both directions, update the index, append a log entry, and commit. A single ingest typically touches 5 to 15 pages.
  • Query (read): to answer a question, read the index, drill into relevant pages, synthesize an answer, and optionally file the answer back as a new page so explorations compound too.
  • Lint (maintain): periodically health-check for orphans, dead links, stale claims, missing frontmatter, and missing cross-references, and apply the corrections.

Why exactly three? Two would conflate adding new knowledge (ingest) with maintaining what is there (lint), and that makes both worse: ingest commits drift from conventions, and lint commits accumulate substantive changes that should have carried their own provenance. Four or more start to overlap. The template's own framing calls three "the empirical minimum." See Everyday-Operations for how to run each.

What does the template capture beyond facts?

Facts are the floor. The richer framing (from the vision wiki's Knowledge, Expertise, Experience page) names three layers a well-curated wiki captures:

  • Knowledge: facts, concepts, definitions, structured information. What most wikis capture by default.
  • Expertise: the author's judgment. Which tradeoffs matter, which approaches to prefer, which criticisms are load-bearing. Captured in synthesis pages, decision records, criticizes:/supports: edges, and "considered and rejected" sections.
  • Experience: the concrete cases behind the abstractions. What was tried, what worked, what failed, edge conditions, dead ends that save future time. Captured in dated log entries with reasoning attached and honest experiment write-ups.

A wiki that records only facts is a reference work; one that records expertise and experience on top is closer to what a graduate advisor provides than to a textbook. See Knowledge-Bundles-Overview and Best-Practices.

What is the difference between knowledge, expertise, and experience in a wiki?

They are separable but not hierarchical (per the vision wiki). Experience can be captured without expertise (raw observation logs); expertise can be asserted without experience (a review of others' work). Each maps loosely onto established literature: knowledge to declarative knowledge and the K of the DIKW pyramid; expertise to procedural knowledge and professional judgment; experience to tacit knowledge and case-based reasoning. The unusual and valuable thing is a wiki with all three deeply captured. See Concepts-Glossary for the individual terms.

Why does the template favor lightweight structural typing over heavy content-graph engineering?

Because heavy content graphs are only as reliable as their entity resolution, and entity resolution is not close to solved for real research corpora. Under-merging ("Obama", "President Obama", "Barack Obama" as three nodes) and over-merging ("Washington" as one node collapsing person, state, and university) both break crisp queries, and they fail worst on exactly the rare, context-disambiguated entities a researcher is most likely to ask about. Structural edges (up: chains, wikilinks, containment) and lightweight extracted content (concept and entity mentions taken at face value) are cheap to build, have no resolution step to fail, and degrade gracefully. The vision wiki argues that for experimental science, lightweight structure walked by a Markov process outperforms heavy typing plus SPARQL on the cost-benefit axis. Aggressive typed-edge discipline still pays off in bibliographic review wikis, where the material is objectively typeable. See Knowledge-Graph.

Why not use SPARQL as the primary retrieval mechanism?

SPARQL over typed edges answers well when the structure being queried is objective and complete; it returns empty or misleading results when the structure is provisional or sparse, which is the common case in experimental science. The template's stance (documented on the KG pages) is that SPARQL is excellent for topology questions ("which pages support this claim?", "what extends this concept?") and for wiki maintenance (orphan detection, hub analysis), but not the right primitive for open-ended question-answering. The recommended pattern is to use the graph to discover where to look, then read the chosen pages to learn what they say. See Knowledge-Graph.

What is the "context envelope" and why does it matter for scientific work?

From the vision wiki's retrieval-context-envelope framing: retrieval is not really about answering a question, it is about assembling the local neighborhood of concepts, layout, and cross-document evidence around the question. When that envelope is missing, the language model cannot answer even when the strictly correct chunk is in the top-K. This matters for science because evidence-based reasoning is multi-source by construction: a claim's support lives in its methods, its dataset, and its replication attempts, scattered across chunks. This is vision-wiki framing rather than a shipped template feature, but it explains why the template invests in a graph substrate rather than dense retrieval alone.

What is a retrieval-hop versus a reasoning-hop?

Also from the vision wiki. A reasoning-hop lives on the question side: decomposing a complex question into sub-questions. A retrieval-hop lives on the index side: the chain of documents you must assemble to build the context envelope. The distinction matters because question decomposition (reasoning-hops) does not by itself assemble the evidence neighborhood (retrieval-hops); dense similarity rank-collapses on the deeper retrieval-hops. This is the empirical motivation for graph-informed retrieval and is documented, with MuSiQue data, on the corpus-side retrieval-context-envelope site.

What is the Matrix helicopter analogy for, and where does it break down?

The vision wiki uses it to name an ambition: in The Matrix, Trinity opens her eyes able to fly a helicopter because she just received a pilot's compiled knowledge, expertise, and experience. The shared destination is rapid competence acquisition, gaining functional competence in a domain fast, at course-load scale, with judgment and cases shipped alongside the facts. Where it breaks down: the film's mechanism is direct brain-writing of skills, which the neuroscience literature places firmly outside the tractable horizon. The template substitutes a different mechanism for the same ambition, the receiver's LLM-assisted workflow rather than their neurology, and unlike the film's opaque transfer, bundle contents are inspectable, extendable, criticizable, and cite-able back to their origin. Used with those qualifications the analogy names the destination without over-promising the mechanism.

What is rapid competence acquisition in the template's framing?

It is the practical goal the Matrix analogy points at: a consumer of a well-curated knowledge bundle, reading it through their own LLM, gains working competence in the bundle's domain far faster than by working through a reading list, because the bundle ships not just the facts but the author's judgment and grounded experience, and the LLM walks that material alongside the consumer's own notes. The Matrix names the ambition; the template names how to approximate it today.

What intellectual traditions does the template draw on?

Three, per the vision wiki and llm-wiki.md:

  • Extended cognition (Bush's Memex, 1945; Licklider's man-computer symbiosis, 1960; Engelbart, 1962; Clark and Chalmers' extended mind, 1998): cognition legitimately extends beyond the skull into external artifacts the agent reliably uses. A bundle plus a local LLM is exactly such an artifact. Bush's Memex is the closest ancestor: a private, actively curated store with associative trails between documents; the part he could not solve was who does the maintenance, and the LLM handles that.
  • Tacit knowledge and reflective practice (Polanyi, 1966; Schön, 1983; Nonaka and Takeuchi, 1995): the ground of the expertise and experience layers.
  • Deliberate practice and expert performance (Ericsson and Charness, 1994): why skill is gradual and embodied, which is also why the Matrix mechanism stays fiction.

See Concepts-Glossary for definitions of these terms.

See also

Clone this wiki locally