wiki

Use when searching for information across your Claude Code workspace, when compiling new knowledge into the wiki, or when running wiki maintenance. Provides a token-efficient search protocol (index → frontmatter → full read), an ingest protocol for filing session learnings back into the wiki, and a maintenance protocol for running the scorer and fixing issues. Trigger keywords: wiki, search workspace, find skill, knowledge base, obsidian, second brain.

Search Protocol

The wiki uses three-tier progressive disclosure. Always start at Tier 1. Only escalate when needed. Do not read full articles to answer questions that the index or frontmatter can answer.

Tier 1: Index scan Read wiki/index.md. Scan the table by title, tags, and summary. This costs minimal tokens and answers most "do we have X?" questions.

Tier 2: Frontmatter scan Read the target article's first ~15 lines (frontmatter only). The summary and related fields tell you whether this is what you need without reading the full body.

Tier 3: Full read Read the entire article body for workflow details, key concepts, and connections. Only do this when Tier 1 and Tier 2 confirm the article is relevant and you need the specifics.

Fallback If no match in the index: grep wiki/ for keywords, then check .claude/skills/ and .claude/plugins/ directly.

Ingest Protocol

When a session produces reusable knowledge that should be filed into the wiki:

  1. Write the article to the appropriate wiki/ subdirectory (skills/, plugins/, or concepts/).

  2. Follow the frontmatter schema exactly:

    ---
title: Article Title
type: skill | plugin | concept
tags: [tag1, tag2]  # from wiki/tags.md only
source: path/to/source
compiled: YYYY-MM-DD
summary: >
  20-80 words, specific enough to decide relevance from the index.
related:
  - "[[Related Article]]"
---

  3. Use only tags from wiki/tags.md. If a new tag is needed, add it to tags.md with a comment description before using it.

  4. Run index_builder.py to update the index:

    python wiki/scripts/index_builder.py

  5. Append an entry to wiki/log.md:

    ## [YYYY-MM-DD] ingest | Article Title
Created from session work. Source: [describe context].

Maintenance Protocol

Run on demand when the wiki needs a health check.

  1. Run the scorer for full validation:

    python wiki/scripts/scorer_wiki.py --post-all

  2. Read wiki/scripts/score_report.json for failures.

  3. Fix interactively or queue for an overnight analyzer run.

  4. Run hub_detector to check for new concept candidates:

    python wiki/scripts/hub_detector.py

    This emits proposal files to wiki/scripts/hub-proposals/ rather than writing articles directly. Review each proposal and change its status frontmatter to approved or rejected before creating the actual topic article.

Conventions

  • Never modify raw source files in .claude/skills/ or .claude/plugins/ — the wiki is the compiled layer.
  • Wiki articles summarize and cross-reference; they don't duplicate source content.
  • New tags require explicit addition to wiki/tags.md before use.
  • File naming: kebab-case matching the source directory name (e.g., diagnosing-energy-models.md).
  • Wikilinks use title case: [[Diagnosing Energy Models]], not [[diagnosing-energy-models]].