Active · v1.0

LLM Wiki for Amplifier

Persistent knowledge that compounds — not RAG that re-derives.

microsoft/amplifier-bundle-llm-wiki · default-on in amplifier-foundation · May 2026

The shape of the problem

RAG re-derives. Wikis compound.

Every RAG query starts from zero — chunk, embed, retrieve, reason, throw it away. The model never accumulates structured understanding of your sources. Andrej Karpathy's LLM Wiki pattern proposes the inverse: the wiki is the compiled artifact. The LLM owns the bookkeeping. Humans curate.

The RAG cycle

Discover · forget · repeat

Sources stay raw. Every query re-runs the same expensive understanding work. There is no place where knowledge accumulates.

The wiki cycle

Mine · integrate · cite

Raw inputs are mined incrementally into entities. The wiki cross-references itself. Queries read from the compiled layer.

The split of labor

LLM bookkeeps · humans curate

The model maintains schema, links, freshness. People decide what's worth ingesting and review what landed.

What shipped

Five modes. One specialist. Zero ambient cost.

A composable workflow surface: each /wiki-* mode is a focused step. They share orientation but never load it unless one is active.

Modes contributed
  • /wiki-init <description> — design and scaffold a project-specific policy (schema, publish target, viewer)
  • /wiki-ingest <source> — mine entities, integrate, cross-ref, log
  • /wiki-lint — read-only health check: orphans, stale refs, contradictions
  • /wiki-publish — run the project's publish script
  • /wiki-query <question> — read-only Q&A against the compiled wiki, with citations
Agent contributed

wiki-policy-designer

Contributed only by /wiki-init's contributes.agents. Mounts only when that mode is active. Body loads only when delegate is called against it. The actual scaffolder behind /wiki-init.

Karpathy pattern, made concrete

The bundle is a working, ecosystem-integrated implementation of the gist — wired into how Amplifier already handles modes, behaviors, and bundle composition.

The flow

init → ingest → lint → publish → query

A normal cadence: scaffold once, then ingest as sources arrive, lint before shipping, publish on demand, query whenever. No mode is required — each is independently useful.

/wiki-initScaffold once
/wiki-ingestMine sources
/wiki-lintHealth-check
/wiki-publishShip artifact
/wiki-query Read-only — at any time, against the compiled wiki

Linear by default, but the modes are independent: you can lint without publishing, query without ingesting, re-run init to evolve the schema. The cycle is a guide, not a state machine.

What it actually costs

Zero ambient. Cost shows up only in use.

The bundle's persistent footprint is zero because YAML config never enters LLM context — it's processed at runtime to wire up structural discovery, then discarded. Only mode bodies and contributes.context files reach the model, and both are mode-gated. See amplifier-bundle-modes/context/mode-schema-reference.md §9.7 for the mechanism; this bundle is its reference implementation.

Session state Tokens in LLM context
Bundle composed; no wiki mode active 0
Any one wiki mode active mode body + ~500 shared orientation
/wiki-init active and delegating to the agent + agent body (only during the delegate call)

Adoption tax for users who never touch the wiki: nothing.

How to get it

If you have foundation, you already have it.

The bundle is default-on in amplifier-foundation. Any standalone that composes foundation — amplifier-dev, with-anthropic, with-openai, and friends — picks it up transitively. The five /wiki-* modes are already discoverable via /modes. Nothing to install.

If you're on a different base and want the wiki bundle as an overlay (works directly with amplifier-app-cli, or on top of any other bundle that has mode infrastructure):

amplifier bundle add --app git+https://github.com/microsoft/amplifier-bundle-llm-wiki@main#subdirectory=behaviors/llm-wiki.yaml

Either way, bootstrap a new project with a brief:

> /wiki-init <brief>

The init flow delegates to wiki-policy-designer, which interviews you about schema, publish target, and viewer — then scaffolds .wiki/, raw/, the publish script, and the project's AGENTS.md entry.

A worked example

An engineering decision log, in three turns.

One plausible shape — design-review transcripts in, an ADR-style decision wiki out. The same flow applies to research-paper summaries, support runbooks, weekly meeting notes, or any compiled-artifact workflow.

> /wiki-init Capture architecture decisions from design-review meetings → ADR-style records (context, decision, alternatives, consequences) → published as a searchable static HTML site for the team. ⤷ delegate(wiki-policy-designer) • drafted schema: decisions, projects, contributors, tags, related-decisions • publish target: decisions-site.zip (static HTML + search index) • viewer: local SPA on :8888 with symlinked data/docs • wrote .wiki/context/schema.md, .wiki/scripts/{publish,verify}.sh • scaffolded raw/, raw/archive/, decisions-package/ • updated AGENTS.md with the cycle docs > /wiki-ingest raw/2026-05-22-api-versioning-review.md ⤷ mining transcript (52min, 4 participants) • new decision: deprecate v1 endpoints, migrate clients by Q3 • linked 2 prior decisions as superseded-by • patched 4 contributor files (positions taken, dissents recorded) • added 3 .md callouts + 1 INDEX entry • moved transcript → raw/archive/ • all sources[] cite the transcript line ranges > /wiki-lint ⤷ 38 JSONs parse · cross-refs resolve · 2 findings, 0 deferred
Lineage and study

The pattern, traced.

Karpathy proposed the shape; multiple implementations have explored corners of it. The Amplifier bundle is informed by that lineage — the contribution is making it composable, ecosystem-native, and dormant-by-default.

Origingist.github.com/karpathy/442a6bf555914893e9891c11519de94f
Skills-shapedalirezarezvani/claude-skills
Direct studybalukosuri/llm-wiki-karpathy
From codebalukosuri/wiki-from-code
Library formlucasastorian/llmwiki
App formxoai/sage-wiki
Quarto angletobi/qmd
Personal-braingarrytan/gbrain
Agent stackNousResearch/hermes-agent

None of these are direct dependencies. They are reference points — the Amplifier bundle's distinctive contribution is the zero-cost-when-dormant integration, not the wiki workflow itself.

Try it

The shortest path from here.

1 · Check

Run /modes

If you see /wiki-init, /wiki-ingest, /wiki-lint, /wiki-publish, /wiki-query — you already have it. Otherwise, overlay it:
amplifier bundle add --app git+https://github.com/microsoft/amplifier-bundle-llm-wiki@main#subdirectory=behaviors/llm-wiki.yaml

2 · Bootstrap

Run init with a brief

/wiki-init <describe what you want to track and ship>

3 · Read the references

Three landing pads

github.com/microsoft/amplifier-bundle-llm-wiki
amplifier-bundle-modes/context/mode-schema-reference.md §9.7
amplifier/docs/MODULES.md (llm-wiki entry)

If you build a wiki with it, the team would love to see the schema you converge on. The patterns aren't fixed — /wiki-init is designed for any compiled-artifact workflow.

LLM Wiki for Amplifier