AI Agents Skills Documentation

Shared, sanitized skill bodies, settings metadata, and installers for Codex, Claude, DeepSeek, GitHub Copilot, OpenCode, Antigravity CLI, Grok, Kimi Code, and restricted OpenClaw targets. The docs describe a research workstation where multiple agent frontends share one canonical skill repository, while local software such as Python, TeX, optional SageMath, library tools, document parsers, and public database clients are detected as external capabilities instead of being hardcoded.

Latest update: Sep 07, 2026

Start Here

First time here: clone the repository, run commands from the repo root, and use doctor, precheck, and plan before any applied install.
New install: read Installation, then choose a Profile.
Existing agent setup: run a read-only audit with Audit And Migration before installing anything.
OpenClaw integration work: use the OpenClaw Integration Plan before adding migration, hook, or runtime behavior.
OpenClaw install-target work: use the OpenClaw Install Target Plan before adding OpenClaw as a writable target.
Dependency questions: use Dependencies, then check the platform-specific Windows or Linux notes.
Runtime-backed skill questions: use Installation for runtime roots and inventory boundaries, then Dependencies for Docling/OCR and local config notes.
Zotero/Calibre local-library changes: run Local Library Profiles before choosing paths.
Unsure what will be written: read Agent Locations and Verification.
Reusable failure or workflow lesson: use self-improving-agent to log the .learnings/ entry and propose a canonical repo integration plan with install-target and OS coverage before editing shared files.
Research workflow motivation: start with System And Research Workflow Overview and Multi-Agent Examples.
Submission venue selector work: use the Submission Venue Selector Plan before implementing the automated venue-ranking skill.
Course teaching workflows (Canvas, Google Classroom, Classroom50, local roster DB): use Course Management Skills and the course-management install profile.
Provisioning the host-local course-management or VNU eOffice application packages: read External Dependency Provisioning before any apply.

Safety Model

The installer is dry-run first. Real home-directory writes require explicit --apply --real-system. Existing unmanaged files are skipped unless the user chooses --adopt, --backup-replace, or --migrate. Secrets, auth files, session logs, local paper libraries, and runtime caches are outside the managed scope.

Runtime-backed skills use managed helper files under a shared runtime root. Live runtime config, caches, local databases, downloaded papers, bytecode, and secrets stay outside the canonical source tree. The Docling wrapper is local-only by default; OCR.space is available only as an explicit fallback with --ocr-fallback ocrspace --allow-remote-ocr.

Most repository docs are generated from installer/ai_agents_skills/docs.py and manifest data. Generated docs are README.md, each page emitted by generated_doc_texts() under docs/, and mirrored copies under docs/source/. Edit the generator or manifests, run make docs, then use make docs-site when you need to preview the Sphinx site. This index.md page, overview.md, submission-venue-selector-plan.md, course-management.md, and external-dependencies.md are maintained manually as docs-site pages (keep their docs/ copies in sync with docs/source/).

Contents