Overview

AI Agents Skills is an experimental, personal-use repository for sharing research-oriented agent skills and settings across Codex, Claude, DeepSeek, OpenCode, Antigravity CLI, Grok, Kimi Code, GitHub Copilot, and restricted OpenClaw. It is designed for combinatorics and graph theory workflows, but the installer and documentation are written so other users can inspect, dry-run, and install only the parts that fit their own machines.

The repository is a generator and installer, not a copied dotfiles folder. It keeps reusable skill bodies, dependency metadata, profiles, optional artifacts, and target-specific rendering logic in one source tree. Agent homes such as ~/.codex, ~/.claude, ~/.deepseek, ~/.config/opencode, ~/.grok, and ~/.kimi-code are runtime targets. Default skill installs use auto mode: Claude links back to the canonical repo files; Codex, OpenCode, Grok, and Kimi receive copied native skill files plus support files; and DeepSeek receives reference adapters unless native loader evidence justifies a different policy. Explicit symlink, reference, and copy modes are available when you need to force one strategy.

Most checked-in documentation is generated from installer/ai_agents_skills/docs.py, with manifest-derived tables inserted from manifest/. Maintainers should edit the generator or manifests and run make docs rather than hand-editing generated README.md or mirrored docs/*.md pages.

Main Ideas

  • Canonical skills live under canonical/skills/. Runtime helpers live under canonical/runtime/. Agent homes and ~/.openclaw/workspace/skills/* are install products — edit the checkout first, then install or publish.

  • OpenClaw dual-route /aas for remote-bridge is published from canonical/runtime/skills/remote-bridge/ via publish_openclaw_adapter.py into ~/.openclaw/workspace/skills/aas-remote-bridge/ (not a managed openclaw-target-* skill-file install).

  • Profiles in manifest/profiles.yaml select useful skill bundles.

  • Optional artifacts add templates, personas, instruction docs, entrypoint aliases, and management notices outside normal skill directories.

  • Runtime-backed skills install helper scripts under a shared runtime root; live config, caches, local databases, and downloaded documents are outside managed canonical source.

  • self-improving-agent turns reusable failures, corrections, and missing capabilities into .learnings/ entries plus canonical repo integration plans that name affected targets, OS/substrates, docs, manifests, runtime helpers, and tests before implementation.

  • precheck detects tools and Python packages from the current substrate.

  • plan and install --dry-run preview writes before anything is changed.

  • --install-mode auto is the default and resolves per agent. Claude uses symlinked skill files. Codex copies complete skill trees because current Codex discovery ignores file-symlinked user SKILL.md files and installed skills must not depend on the source checkout. DeepSeek uses reference adapters because native symlinked skill loading has not been verified, and OpenCode, Grok, and Kimi use copied native skill files. symlink, reference, and copy force one strategy for every agent.

  • Real home-directory writes require explicit --apply --real-system.

  • Verification checks only installed managed artifacts.

  • The Docling document/OCR runtime is local-only by default. Stronger scanned PDF extraction uses local presets such as scan-heavy; OCR.space is only an explicit opt-in fallback when local conversion fails or quality degrades.

Typical Workflow

Clone the repository and run commands from its root before starting:

git clone https://github.com/hoanganhduc/ai-agents-skills.git
cd ai-agents-skills

make doctor
make precheck ARGS="--profile research-core"
make audit-system ARGS="--profile research-core"
make plan ARGS="--profile research-core"
make install ARGS="--profile research-core --dry-run"

Use a fake root for write testing:

make lifecycle-test ARGS="--matrix default --platform-shape linux"
make verify ARGS="--root <fake-or-real-root>"

Where To Go Next

  • Installation: safe install, dry-run, conflict, and migration flows.

  • Skills: canonical skill catalog.

  • Profiles: workflow bundles such as research-core, library, math, full-research, and course-management.

  • Course Management Skills: Classroom50, Canvas, Google Classroom, and local roster DB agent entrypoints.

  • Optional Artifacts: templates, personas, instruction docs, entrypoints, and management notices.

  • Dependencies: logical tools, Python packages, and platform-specific detection behavior.

  • Workflow Overview: how the research stack connects agents, skills, runtimes, and external software.

  • Audit And Migration: how to inspect an existing setup before adopting or migrating files.

  • OpenClaw Install Target Plan: how to stage future OpenClaw writable target support behind evidence gates.

  • Verification: what installed managed artifacts are checked after install, uninstall, or rollback.