LLM Wiki and UAIX Project Handoff

How to read this report

This is a dated report about the boundary between LLM Wiki knowledge accumulation and UAIX Project Handoff repository-state transfer. Use it as explanatory context after AI Memory and Project Handoff.

LLM Wiki is treated as an adjacent knowledge-base pattern, not a UAIX support surface.
AI Memory and Project Handoff remain repository-context guidance; public exchange and release claims still need UAI-1, Validator, release notes, or implementation evidence.
Mentions of generators, upload validators, SDKs, CLIs, certification, or endorsement remain future work unless the public Roadmap, Changelog, and published tooling evidence say otherwise.

LLM Wiki and UAIX Project Handoff: A Layered Model for Durable AI Context

The practical question is not whether to use LLM Wiki or Project Handoff. It is where each one belongs, who is allowed to edit it, and what kind of claim it can support.

Publication boundary

This article treats Project Handoff as the current UAIX repository-context draft and treats LLM Wiki as an adjacent knowledge-management pattern. It does not claim that UAIX currently provides a public .uai generator, upload validator, SDK, CLI, certification, or endorsement program.

Why The Two Should Be Combined

Karpathy’s LLM Wiki pattern is strong because it gives AI a durable place to compile what it has learned from sources. UAIX Project Handoff is strong because it gives the next AI a durable place to find the current working state of a repository. The two become much more useful when they are combined with a clear boundary: the wiki can inform the handoff, but the handoff governs execution.¹ ⁶

That boundary mirrors the way OpenAI’s Codex documentation frames project guidance. AGENTS.md is durable project guidance that travels with the repository, while memories, skills, MCP, and other layers serve different jobs. UAIX adds a stricter project-handoff convention by splitting stable context into typed .uai files and requiring a first response before broad edits.⁴

Recommended Repository Shape

A combined repository should keep long-running synthesis and operational truth physically separate. A compact layout looks like this:

AGENTS.md
.uai/
  context.uai
  stack.uai
  constraints.uai
  architecture.uai
  decisions.uai
  progress.uai
  operations.uai
wiki/
  index.md
  log.md
  sources/
  concepts/
  comparisons/

The wiki/ folder is where the LLM builds accumulated knowledge from sources. The .uai/ folder is where the project records what is current, what is constrained, and what the next actor should do. The root AGENTS.md should state that hierarchy plainly.

The Promotion Rule

Do not let generated wiki text become operational instruction automatically. Promote an idea from the wiki into .uai only after a human or maintainer decides it is current project truth. A useful promotion path is:

Capture raw source material in the wiki source layer.
Let the LLM summarize, compare, and cross-link the material in the wiki.
Review the synthesis against source links before relying on it.
Move only approved operational facts into the relevant .uai file.
Update AGENTS.md when the load order, current state, or hard constraints change.
Use UAI-1 and validator evidence when the work becomes an external exchange or public support claim.

This keeps the wiki useful without letting a generated concept page quietly rewrite the rules of the repository.

What Goes Where

Material	Put it in LLM Wiki	Put it in Project Handoff
Research notes	Yes, especially source summaries and comparisons.	Only if they become current project facts.
Build and test commands	No, except as background.	Yes, usually in `operations.uai` or `AGENTS.md`.
Security boundaries	Can explain the rationale.	Yes, as hard constraints in `constraints.uai`.
Architecture history	Useful for narrative context.	Use `architecture.uai` and `decisions.uai` for current accepted decisions.
Known failed attempts	Useful as research trail.	Use `errors.uai` or `progress.uai` when the next AI must avoid repeating them.
Public interoperability evidence	No.	No by itself. Use UAI-1, validator output, conformance packets, changelog, and implementation records.

The First-Response Pattern

A mature combined setup still needs a moment of explicit confirmation before action. UAIX Project Handoff asks the next AI to summarize the project, list loaded context files, confirm hard constraints, name intended touchpoints, and state checks before broad changes. That pattern is especially important when a wiki is present, because it forces the agent to say which material it is treating as binding project context and which material is only background.⁶

Public Claims Need More Than Memory

Neither a wiki page nor a handoff file is public conformance evidence. They can prepare the work, but they do not prove an exchange, package, or implementation claim. When the output needs to travel beyond the repository, the evidence path should move into UAI-1, published schemas, registry records, examples, validator behavior, conformance packets, roadmap boundaries, and changelog entries.⁸ ¹⁰

That is the simplest way to keep the layers honest: LLM Wiki accumulates knowledge, Project Handoff transfers project state, and UAI-1 carries portable public exchange evidence.

Editorial Guidance For UAIX

UAIX should present the two patterns as complementary rather than competitive. The clean positioning is:

LLM Wiki helps AI-assisted knowledge accumulate. UAIX Project Handoff helps AI-assisted project state transfer.

That sentence keeps the comparison easy for readers while preserving support boundaries. It lets UAIX cite the LLM Wiki conversation without implying that Project Handoff is a general-purpose knowledge base, and it lets UAIX explain Project Handoff without claiming that local files are already UAI-1 conformance records.

Works Cited

Andrej Karpathy, LLM Wiki, GitHub Gist, created April 4, 2026.
AGENTS.md, A simple, open format for guiding coding agents.
OpenAI Developers, Custom instructions with AGENTS.md – Codex.
OpenAI Developers, Customization – Codex.
Patrick Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, arXiv:2005.11401.
UAIX, Project Handoff.
UAIX, AGENTS.md .uai Linking Specification.
UAIX, UAI-1.
UAIX, Standards Fit.
UAIX, Validator.