AI-Agent Work Wiki

Obsidian knowledge base with raw/wiki/conversations tiers and fixed commands so multiple AI agents work consistently.

Period 2026.06 – present · Type Personal project · Role Design and operation

Why I Built It

Robotics R&D mixes activities such as paper searching, source code analysis, report and paper writing, projects, and job-search preparation, and the more AI agents are used, the bigger the problem of context being lost between sessions becomes. What was needed was not a personal notepad but a work system that is easy for a person to read and edit while multiple agents execute it consistently.

Design

  • Three-tier separation (Karpathy’s LLM Wiki principle): raw/ (originals, immutable) → wiki/ (verified knowledge only) ← conversations/ (temporary, handover). Because crossing these boundaries is the main cause of context contamination, the promotion flow is made explicit.
  • Personal/company hybrid: Shared knowledge lives in one tree with a context: personal | company tag; projects and documents are split into separate folders.
  • Fixed command keywords: save (store after passing a five-filter gate) · reference (restore context) · ingest (process originals into the wiki) · lint (audit structure, links, speculation, duplicates, sources, and security). Each maps 1:1 to a natural-language trigger (“save this session’s work to Obsidian”).
  • Safeguards against context contamination: Source principle, no speculation, no duplicates, and reversals handled as superseded. Saving to the wiki requires satisfying at least one of five filters: reuse, handover, decision traceability, failure risk, or shared rules.
  • Dedicated agents: A lab notebook organizer, a paper reviewer, and others run as subagents with fixed output formats. The paper reviews and project pages on this site were also generated from this wiki.

What I Learned

From an experience of mistaking examples for real work and over-specializing before the domain was settled, then having to roll it back, I derived the rule “no specialization before the domain is fixed.” Like a robot system, an agent system only scales when the interface conventions are fixed first.

Tech Stack

Obsidian · Claude Code · Markdown / YAML front matter · Subagents · MCP