Wiki Author — Long-form Documentation Layer

Keywords

adr, spec, decision, design doc, rfc, file doc, notes, wiki, documentation, link pages, bidirectional link, supersedes, implements, depends on

Overview

Cortex's wiki is a Markdown authoring surface for the long-form artifacts that don't fit the thermodynamic memory model: architecture decision records, specs, per-file documentation, and free-form notes. Pages live under ~/.claude/methodology/wiki/ and are never pruned — they are first-class authored content, not derived views. Every write also registers a protected pointer memory in PostgreSQL so recall surfaces wiki pages alongside regular memories.

Use this skill when: the user is making a decision that should persist, finalizing a spec, documenting a file's purpose, or asking for two wiki pages to be linked.

Do NOT use for: ephemeral facts (use remember), domain profiles (use query_methodology), or regenerating documentation from memory (the wiki is authored, not projected).

Workflow

Record an architecture decision

cortex:wiki_adr({
  "title": "Use pgvector for retrieval",
  "context": "We need a searchable memory store with hybrid ranking.",
  "decision": "Adopt PostgreSQL + pgvector + pg_trgm as the single backend.",
  "consequences": "No SQLite fallback. Operator must provision Postgres 15+.",
  "status": "accepted",
  "tags": ["storage", "retrieval"]
})

Returns { path, number, title, status, ... }. ADR numbers auto-increment.

Write a spec

cortex:wiki_write({
  "path": "specs/wiki-authoring.md",
  "content": "# Wiki authoring\n\n## Summary\n\nClaude authors ADRs and specs during sessions..."
})

Pass the final markdown you want to land. Use "mode": "append" to add to an existing page, "replace" to overwrite.

Document a file

cortex:wiki_write({
  "path": "files/mcp_server-handlers-wiki_write.md",
  "content": "# `mcp_server/handlers/wiki_write.py`\n\n## Purpose\n\nComposition root for the wiki authoring tool..."
})

Link two pages bidirectionally

cortex:wiki_link({
  "from_path": "adr/0001-use-pgvector-for-retrieval.md",
  "to_path": "specs/retrieval-pipeline.md",
  "relation": "implements"
})

Adds the forward relation to from_path and the inverse (implemented_by) to to_path. Idempotent. Known relations: supersedes/superseded_by, implements/implemented_by, depends_on/depended_on_by, derived_from/derives, see_also.

Read a page (to update it)

cortex:wiki_read({ "path": "adr/0001-use-pgvector-for-retrieval.md" })

List pages

cortex:wiki_list({})                   // all kinds
cortex:wiki_list({ "kind": "adr" })    // just ADRs

Regenerate the table of contents

cortex:wiki_reindex({})

Writes .generated/INDEX.md. This is the only file ever auto-regenerated — authored pages are untouched.

Layout

~/.claude/methodology/wiki/
  adr/NNNN-<slug>.md       # numbered decision records
  specs/<slug>.md          # feature specs, PRDs, design docs
  files/<slug>.md          # per-source-file documentation
  notes/<slug>.md          # free-form notes and investigations
  .generated/INDEX.md      # auto-generated table of contents

Tips

• No prune, ever. Authored pages survive across sessions, version bumps, and machine moves (as long as the wiki root is preserved).
• ADR numbers are monotonic. Don't reuse them — supersede instead via wiki_link(new, old, "supersedes").
• Links live in the page. The ## Related section is the source of truth for wiki-internal links — not a sidecar index.
• Wiki pages are recallable. After wiki_write, the page content (first 500 chars) is indexed in PostgreSQL as a protected memory tagged wiki, so recall returns it alongside regular memories.
• Back-links are authoritative. wiki_link always updates both pages. If you move or rename a file, re-run wiki_link so the related sections stay consistent.