Show HN: A Karpathy-style LLM wiki your agents maintain (Markdown and Git)

Apr 25, 2026•najmuzzaman•View Original

TL;DR Highlight

WUPHF builds a shared knowledge base using a Git-based Markdown Wiki, enabling multiple AI agents—including Claude and Codex—to autonomously divide and execute tasks.

Who Should Read

Backend and full-stack developers orchestrating multiple AI agents to build automation pipelines, or developers experimenting with long-running agent workflows using Claude Code/Codex CLI.

Core Mechanics

WUPHF structures AI agents—like Claude Code or Codex CLI—as 'AI employees' with roles such as CEO, PM, and engineer, allowing them to autonomously claim and perform tasks within a shared office environment.
To prevent context loss, agents rely on a Markdown-based Wiki managed with Git as a 'shared brain,' applying Karpathy’s proposed wiki style to a real-world multi-agent setting.
The UI supports two modes: a `--tui` mode based on a tmux terminal UI and a web UI for running agents headlessly, requiring Claude Code or Codex CLI and tmux as prerequisites.
Search employs a BM25-first routing approach, using a heuristic classifier to route short, simple queries to BM25 and more descriptive queries to a cited-answer loop, reporting 85% recall@20 on 500 artifacts.
The system aims for a positive feedback loop where agents document generated knowledge in the Wiki, which then serves as context for subsequent agent runs.
Directory structure—including `prompts/`, `evals/`, `mcp/`, and `claude-code-plugin/commands/`—reveals a full-stack agent infrastructure encompassing prompt management, evaluation, MCP tool integration, and Claude Code plugins.
AGENTS.md, ARCHITECTURE.md, DESIGN-WIKI.md, and TESTING-WIKI.md are included in the repository, enabling agents to read and understand contribution rules.

Evidence

"Skepticism arose regarding the value of AI automatically accumulating notes, as critical reading and integration into one’s mental model are core to note-taking—simply increasing document volume misses the point."

How to Apply

If multiple Claude Code sessions struggle with context initialization when handling the same codebase, leverage WUPHF’s AGENTS.md + DESIGN-WIKI.md pattern by designating a repository root as a knowledge document for agents and specifying its inclusion in system prompts at session start.
To mitigate quality degradation in automatically generated Wiki content, divide it into 'draft' and 'confirmed' directories, injecting only files from the confirmed folder as agent context to filter noise.
When implementing a BM25 + cited-answer hybrid search, adding explicit rules—such as routing queries containing specific keywords (e.g., function names, file path patterns) to BM25 and others to semantic search—improves stability over relying solely on query length or sentence structure.
If long-running agents frequently fail and waste tokens, break tasks into 5–10 minute checkpoints, commit results to the Wiki after each, and initiate subsequent steps with new agent calls to reduce cumulative failure probability.

Terminology

BM25A classic text search algorithm considering keyword frequency and document length. Faster and simpler than vector search, excelling at direct word matching.

cited-answer loopA RAG (Retrieval-Augmented Generation) style processing flow generating answers while citing sources. More costly and slower than simple keyword search but suitable for queries requiring contextual understanding.

recall@20The proportion of times the correct answer is within the top 20 results. 85% recall@20 means the desired document appears in the top 20 results 85 out of 100 searches.

MCPShort for Model Context Protocol, a standardized protocol enabling AI models to access external tools and data sources, spearheaded by Anthropic.

headlessRunning without a graphical user interface (UI). Used for running agents in server or CI environments lacking a GUI.

agent driftThe phenomenon where an agent’s behavior gradually deviates from its initial goals or accumulates unexpected actions over extended runtime.