LLM Wiki

Appearance

Built 26/04/19 07:55commit 447639a

Codex LLM Wiki Skill ​

中文 | English

Summary ​

This source operationalizes the abstract LLM Wiki pattern as a reusable Codex skill bundle: it packages the maintainer posture, root-resolution order, bilingual rules, scheduler model, PDF-ingest policy, and helper scripts needed to run a repository-backed markdown wiki as an ongoing system rather than as a loose prompt pattern.

Source ​

Translation Freshness ​

  • The imported raw pair now reflects the current skill rule set, including the newer PDF ingest path, narrower taxonomy-patrol guidance, and explicit pre-push local verification step.
  • The bundle's raw markdown entry points now expose current bilingual navigation across the maintainer spec and its reference documents.

Key Contributions ​

  • Turns the LLM Wiki idea from a conceptual gist into an executable maintainer method for Codex.
  • Makes root resolution explicit, so the same skill can operate across many vaults without hard-coded repo assumptions.
  • Defines a bilingual maintenance discipline that treats .zh.md files as maintained siblings instead of one-off exports.
  • Splits automation into content patrols and taxonomy patrols, which keeps ingest drift and directory refactors on separate cadences.
  • Promotes raw PDF backlog handling into the default maintenance loop: preserve the original PDF, create a markdown raw sibling, choose short-form versus long-form translation paths, and keep visual evidence when layout matters.
  • Packages helper scripts around narrow structural tasks: vault bootstrap, OCR extraction, linting, scheduled runs, scheduler state/config plumbing, safe path moves for both raw/ and wiki/, and git-remote sync helpers.

Strongest Claims ​

  • The model, not the scripts, should still do the reading, synthesis, contradiction handling, and page revision.
  • index.md and log.md are operating primitives, not incidental files.
  • Raw source substance stays stable, but raw paths may be reorganized when family structure or navigability demands it.
  • Translation freshness is an ongoing maintenance obligation; stale siblings should be marked explicitly instead of silently drifting.
  • Automation should continue from prior state and avoid overlapping runs rather than blindly restarting each patrol.
  • Scheduled or manual maintenance that changes rendered output should finish with a local verification or build step before commit and push.
  • Routine cleanup is not just hygiene: it should first check whether one or two narrow, already-obvious taxonomy moves would improve navigability more than cache or ignore-file cleanup.

How It Changes This Vault ​

  • Adds concrete script-backed implementations for the maintenance loops previously described only at the concept level.
  • Sharpens the distinction between abstract pattern pages and reusable operational bundles.
  • Makes llm-wiki itself legible as a source family: a bundle composed of one maintainer spec, three reference docs, and a small but now fully enumerated tool surface.
  • Gives this vault an explicit policy for future PDF backlog work instead of leaving PDF ingest quality and translation strategy implicit.

Open Questions ​

  • Whether this vault should eventually ingest the helper scripts individually as first-class implementation sources rather than as one bundled operational source family.