knowledge_management

ai-modules

A collection of professional AI skills, agents, commands, and hooks. The same source of truth ships through several equal paths. You can install it from the bundled Claude Code marketplace, symlink it into vendor config dirs globally with make deploy (VS Code Copilot, Cursor, Claude Code, OpenAI Codex, Gemini CLI, Google Antigravity), symlink it into a single repo's local config via --project-dir, or use it in-place from a checkout. The deployment script discovers artefacts by plugin layout and installs them where each tool expects them.

Layout

ai-modules/
├── .claude-plugin/
│   └── marketplace.json     # registers the plugins below as a Claude marketplace
├── plugins/                 # one subdirectory per plugin
│   ├── knowledge_management/
│   │   ├── .claude-plugin/plugin.json
│   │   ├── .codex-plugin/plugin.json
│   │   ├── README.md
│   │   ├── agents/          # one .md file per agent
│   │   │   └── wiki_auto_shaper.md
│   │   └── skills/          # one subdirectory per skill, each with SKILL.md
│   │       ├── wiki/
│   │       ├── wiki_wrapup/
│   │       ├── wiki_import/
│   │       ├── wiki_fix/
│   │       ├── executive_summary/
│   │       └── spr/
│   └── ai_dev/
│       ├── .claude-plugin/plugin.json
│       ├── .codex-plugin/plugin.json
│       ├── README.md
│       └── skills/
│           ├── git_commit/
│           ├── update_changelog/
│           ├── ai_instruction_writing/
│           ├── ai_instruction_formatting/
│           ├── format_markdown/
│           ├── format_python/
│           └── format_rust/
└── deployment/              # deployment script for installing artefacts globally or per-project
    ├── deployment.sh
    ├── deployment.conf
    └── README.md

Plugins

Each skill is a written procedure the model loads when its trigger fires. Bundling deterministic helpers (bash and python scripts, linters, schema files) alongside the prose lets the agent offload mechanical work to programs that can't hallucinate, and follow a written workflow instead of re-deriving one each session. The practical effect is fewer turns per task, smaller context per turn, and more consistent output across runs. On metered models, that translates directly into time and tokens saved.

knowledge_management

Skills and agents for building, maintaining, and distilling a persistent, compounding knowledge base. Everything is plain markdown, readable in any editor or CLI, with no Obsidian or vendor reader required.

The wiki itself plus paired front ends that wrap two of its workflows so the model has a single named entry point per use case:

• wiki: the foundation. Builds and maintains an interlinked markdown wiki — ingest URLs, articles, papers, PDFs, transcripts, meeting notes, internal notes, and pastes; query, lint, audit, archive, and reorganise. The page-type enum (entity, concept, comparison, summary, query, procedure) is read from SCHEMA.md, so a wiki extends its taxonomy without touching the linter. Provenance is anchored by footnotes plus body-only sha256 drift detection on raw sources. Discovery, init, lint, and the sha256 helper all ship as bundled scripts, so the agent runs deterministic programs for the mechanical parts instead of inventing them inline each session.
• wiki_import and wiki_wrapup: triage-first ingest pair. wiki_import takes one named resource (URL, file, paper, PDF, transcript, meeting note, internal note, or paste); wiki_wrapup takes the current chat session. Both capture the source, diff each candidate against the existing wiki, and emit a triage report (new pages, extensions, contradictions with both excerpts and concrete reconciliation options) before any wiki-page write lands. Approved writes route back through the wiki skill. Use them when "review what you'd change before changing it" matters — for example, after a research chat, or before importing a contested paper.
• wiki_fix and wiki_auto_shaper (agent): paired audit-and-repair. wiki_fix is the one-shot skill wrapper; wiki_auto_shaper is the agent it hands off to. The agent runs a two-phase loop — assess (lint plus semantic audit), then fix, then re-lint until clean. It repairs frontmatter and schema violations, broken links, off-taxonomy tags, oversized or topic-mixing pages, procedure pages leaking instance content, content that drifts from the per-type page anatomy, and surfaces cross-page contradictions via the contested-page protocol for human review rather than auto-resolving them.

Two distillation skills that operate on text outside the wiki: