LLM Wiki Evolution

English | 中文

🆕 2026-09-09 — New practice output: The Seven-Layer Tower (《七层塔》), a fact-grounded nonfiction novel about how the AI agent harness stack evolved — L1 Prompt → L2 Tools → L3 Memory → L4 Skills → L5 Subagents → L6 Evaluation → L7 Self-Improvement — written entirely on top of a production wiki maintained with this skill. Read it as full article (HTML) · PDF · interactive edition.

A skill that evolves and governs a Karpathy-style LLM wiki into a reliable decision-grade knowledge system. Use it to audit an existing wiki, fix structural issues, ingest sources with automatic backlinking, and enforce measurable quality gates.

What it does

• Audits wiki health (unresolved wikilinks, orphan pages, duplicate concepts, stale evidence) with an Obsidian-style resolution model.
• Governs naming with a canonical-name policy and alias-to-canonical mapping.
• Auto-links sources: scans an ingested source page for concepts that already exist in the wiki and appends a "Related sources" backlink section — idempotently, and dry-run by default.
• Creates alias stubs for missing concepts, either one-off or in bulk from a health report.
• Adds evidence and decision layers so high-impact claims carry source links, verification dates, and ADR-style decision records.
• Enforces quality gates that fail maintenance runs when unresolved links rise or evidence coverage drops below threshold.

Layout

SKILL.md                              # main skill entrypoint
agents/openai.yaml                    # Codex/agent interface metadata
references/
  canonical-naming.md                 # naming policy and mapping format
  governance-checklist.md             # quality gates, evidence & ADR rules
scripts/
  ingest_url_to_wiki.js               # single entrypoint URL -> wiki page
  wiki_health_check.py                # v1 health audit + canonical CSV starter
  wiki_health_v2.py                   # v2 health scanner (Obsidian-style resolution)
  auto_link_source.py                 # append "Related sources" backlinks to a source page
  make_alias_stub.py                  # generate alias stub pages (single or bulk)

Quick start

# Audit a wiki (v2: Obsidian-style link resolution, excludes backups/snapshots)
python scripts/wiki_health_v2.py --root <wiki_root>

# v1 audit + generate a canonical-map.csv starter from unresolved links
python scripts/wiki_health_check.py --root <wiki_root> --emit-canonical-template

# Ingest a URL into the wiki
node scripts/ingest_url_to_wiki.js --url "<url>" --wiki-root "<wiki_root>" --tags "#ingest #source"

# Auto-link an ingested source page (dry-run first, then --apply)
python scripts/auto_link_source.py --wiki <wiki_root> --source sources/<page>.md
python scripts/auto_link_source.py --wiki <wiki_root> --source sources/<page>.md --apply

# Create one alias stub pointing at its best source
python scripts/make_alias_stub.py --root <wiki_root> --name "My_Concept" \
  --target sources/2026-01-01-some-source.md

# Bulk-create stubs for broken references seen >= 2 times in a health report
python scripts/make_alias_stub.py --root <wiki_root> \
  --from-health wiki-health-v2.json --min-freq 2

The daily ingest pipeline

Every ingested article goes through this exact loop on the production wiki — classify, patch idempotently, auto-link, then verify against a running baseline. GitHub renders the diagram natively:

flowchart TD
    A["WeChat article URL"] --> B["ingest_url_to_wiki.js"]
    B --> C["sources/YYYY-MM-DD-slug.md<br/>(draft + frontmatter)"]
    C --> D{"Human read &<br/>classify"}
    D -->|"substantive concept,<br/>2nd independent mention"| E["Grow existing concept page<br/>(create page only on two-strike)"]
    D -->|"person / product / tool"| F["Entity page section"]
    D -->|"opinion piece / launch note /<br/>course promo"| G["Light touch: one entity line +<br/>source tagging only"]
    E --> H["build_*.py idempotent patch<br/>UTF-8-sig · only link pages that exist"]
    F --> H
    G --> H
    H --> I["auto_link_source.py --apply<br/>LLM extraction: matched backlinks<br/>+ unmatched-concept backlog"]
    I --> J["wiki_health_v2.py --root"]
    J --> K{"Diff vs baseline:<br/>pages / wikilinks /<br/>unresolved / orphans"}
    K -->|"unresolved rises"| L["Find & fix the new dead link,<br/>re-verify"]
    L --> J
    K -->|"flat"| M["Ship · unmatched concepts stay<br/>as the page-creation backlog"]

Loading

1. Ingest — ingest_url_to_wiki.js lands a draft source page with frontmatter (source_url, captured_at, tags).
2. Classify (human pass) — substantive concepts grow existing pages; a brand-new concept page is created only after two independent sources mention it substantively (the two-strike rule). People, products and tools belong on entity pages. Opinion pieces, launch notes and course promos get a light touch only: one entity line plus source tagging.
3. Patch — a one-off build_*.py applies the enrichment; it skips when its section marker already exists (idempotent), writes UTF-8-sig, and may only link pages that already exist.
4. Auto-link — auto_link_source.py extracts concepts with an LLM, appends the "Related sources" section, adds backlinks to matched pages, and records unmatched concepts on the source page — that block doubles as the page-creation backlog queue.
5. Verify — wiki_health_v2.py diffs the four metrics against the running baseline. unresolved must stay flat: any rise means a new dead link was introduced, and the run does not ship until it is fixed.

Hard-won practices from production

These rules were baked into the scripts after running this workflow daily on a 1300+ page production wiki.

Idempotent linking. auto_link_source.py skips source pages that already carry an auto-linked "Related sources" section, so re-running after manual edits never duplicates or clobbers annotations. It defaults to dry-run; --apply is the explicit mutation gate.

Snapshot hygiene in health stats. wiki_health_v2.py excludes .bak files, backup directories, and log snapshots from live-link counting. Without this, stale copies "launder" orphan and dead-link numbers and mask real regressions.

Three-part link normalization. The link resolver treats page keys as equal after (1) stripping the .md suffix, (2) stripping hyphens, and (3) lowercasing and removing spaces/underscores. Wikis mixing [[Agent_Memory]] with sources/agent-memory links produce exploding false positives without all three normalizations.

Second-mention page creation ("two-strike rule"). A concept earns its own page only after two independent sources reference it substantively. A single mention, or a weak/passing reference, goes into an existing page or stays as an unmatched-concept note in the source page — that note block is the backlog queue for future page creation.

Zero-dead-link gate. New pages may only link to pages that already exist. Combined with the health scanner as a post-ingest check, this keeps unresolved flat while the wiki grows. A rising unresolved count fails the maintenance run.

Smoke-test before publishing tool changes. After modifying any script, run it against the real wiki and diff the health numbers against the known baseline before shipping. The scanner's --root parameter exists precisely so the packaged copy can be validated against a live wiki without hard-coded paths.

License

MIT (see LICENSE).