Wiki Content Program Policy
Agreed 2026-08-08 after a full decision-tree review; this page is the durable, harness-independent record. Any agent doing content work on this wiki follows these rules unless the operator changes them.
The nine decisions
- Goals. Agent retrieval quality (A) and personal working reference (C) are jointly primary. SEO (B) is pursued explicitly but only via moves that cost A/C nothing — search-phrased titles as tiebreakers yes, keyword-stuffed summaries no.
- Taxonomy. Targeted cleanup only, never big-bang re-clustering. (Executed: quant-finance split from machine-learning; testing pages consolidated into software-engineering-practices.)
- Topic selection. Periodic (quarterly-ish) domain-completeness seeding passes; between them, the demand loop (real query misses) and project-anchored writing drive what gets written. Syllabus-completeness alone is not a reason to write a page.
- Format follows shape. Comparative/conceptual →
type: article; procedural → type: runbook (requires the structured runbook: frontmatter block; emits HowTo JSON-LD). Never dual-track one topic into both — sibling pages compete in retrieval. - Accuracy tiers. Fast-decaying claims (tool versions, licensing, maintenance status, API spellings) get web-verified and the page stamped via
mark_page_verified; slow-decaying pattern content rides on the annual staleness check. - Knowledge Graph. Practice/context clusters stay default-excluded from KG extraction until a concrete need appears.
- Cadence mechanism. The recurring checks live in the wiki-audit skill: under-served-query sweep + tool-practice staleness flag — see Wiki Audit Demand Loop.
- SEO within the free constraint. Structured-data spot checks, marketing-site inbound links (allowlist convention), search-phrasing as a titling tiebreaker. Nothing beyond.
- Workflow. Publish-direct (no draft gate), tranches of 6–10 pages so per-page verification stays honest, retrieval levers per the wiki-content skill (summary/title/cluster/headings are embedded in every chunk).
The golden rules inherited from measurement
- A clean
verify_pages lint is necessary, not sufficient — the live assemble_bundle check is the real retrievability gate. - Every content edit can help or hurt corpus-wide recall — enrich with discriminating vocabulary, never generic topic lists (a mechanical enrichment pass over 120 pages once dropped section recall 0.74→0.68).
- Dead levers, measured and rejected — do not re-attempt without new evidence: rerankers, HyDE, doc2query, KG graph rerank (deleted from the code in 2026-07 after showing zero lift), lexical injection.
See Also