Personal Agent Skills

Published source of truth for my global agent instructions, personal agent skills, curated third-party skill list, and related agent tooling references.

This repo has four jobs:

1. Own the global AGENTS.md guidance I want agents to inherit.
2. Own and publish skills I write in skills/.
3. Track the curated third-party skills I use in curated-skills.json.
4. Track non-skill tools and references I use in curated-tools.json.

Installs and updates use the open skills CLI from Vercel. The local curated installer is a thin wrapper around curated-skills.json and npx skills add:

npx skills --help

Layout

.
├── skills/                  # Personal skills authored in this repo
├── AGENTS.md                # Global agent instruction source of truth
├── curated-skills.json      # Third-party skills I use, with sources and selected skill names
├── curated-tools.json       # Non-skill tools, packages, apps, and reference docs I use
├── docs/skill-inventory.md  # Full curated skill/plugin/MCP inventory
├── schemas/                 # JSON schemas for repo metadata
├── .claude-plugin/          # Claude Code plugin metadata
├── .codex-plugin/           # Codex plugin metadata
├── .cursor-plugin/          # Cursor plugin metadata
└── package.json             # Convenience scripts around npx skills

Install

Install the published personal skills globally for Cursor, Claude Code, and Codex:

npx skills add luabagg/agent-skills --global --agent cursor --agent claude-code --agent codex --skill '*'

That command installs only the personal skills stored in this repo's skills/ directory. Curated third-party skills are references in curated-skills.json, so they must be installed from their upstream sources.

Without cloning this repo, install curated skills-cli entries by reading curated-skills.json and running npx skills add for sources where preferredInstall is skills-cli. Plugin-style entries are not installable through this repo's GitHub command; install those through each agent's plugin or connector marketplace.

For local development from a checkout, install from the working tree instead:

npm run install:personal

If the skills CLI installs into ~/.agents/skills but Codex does not pick those skills up, mirror the active skill tree into ~/.codex/skills:

npm run sync:codex-skills

The sync preserves Codex's .system skills, skips archives and dot directories, and refuses to overwrite a real directory that already exists in ~/.codex/skills.

By default, npx skills may install with symlinks so the installed agent paths point back to the source. If an agent or filesystem has trouble with symlinks, copy instead:

npm run install:personal:copy

List the skills exposed by the local checkout:

npm run skills:list

Personal Skills

Add personal skills under skills/<skill-name>/SKILL.md.

Each skill should have the standard frontmatter:

---
name: my-skill
description: What this skill does. Use when ...
---

Current personal skills:

Skill	Purpose
branch-port	Port a feature across heavily diverged branches without unsafe merges
memory-palace	Ingest, query, and lint the personal Obsidian knowledge vault
thorough-pr-review	Review PRs and branches for correctness, reliability, and merge-readiness

Every published skill in skills/ must be public-safe. Do not include private dashboards, tokens, org IDs, internal URLs, generated memory context, or workflow details that only make sense inside a private environment. External IDs or URLs are acceptable only when they are intentionally part of the skill's public purpose or point back to this repo's own public references.

Curated Third-Party Skills

Third-party skills are referenced, not copied, unless I intentionally fork or customize them.

For the full curated map of installed skills, plugin packs, MCP-only tools, and not-installed decisions, see docs/skill-inventory.md.

The curated list currently tracks:

Source	Purpose
obra/superpowers	Development workflow skills: brainstorming, planning, TDD, debugging, reviews, and branch finishing
JuliusBrussee/caveman	Compressed communication, commit, review, and compression skills
kepano/obsidian-skills	Obsidian, Markdown, Bases, Canvas, and Defuddle skills
vercel-labs/skills	find-skills for discovering more skills
emilkowalski/skill	emil-design-eng for UI polish, component design, and animation guidance
kunchenguid/gnhf	gnhf for launching, supervising, and reviewing long-running agent-managed coding runs
cursor-public/sentry	sentry for read-only Sentry issue and event investigation
GitHub plugin skills	gh-address-comments and gh-fix-ci for PR comments and GitHub Actions CI triage
Notion plugin skills	notion-research-documentation for Notion research and documentation
Figma plugin skills	figma and figma-implement-design for Figma MCP and design implementation
Linear plugin skills	linear for Linear issue, project, and team workflows

Install the curated third-party set globally:

npm run install:curated

The install script reads curated-skills.json and installs sources where preferredInstall is skills-cli. Plugin-style skills are tracked as references and should be installed through their agent/plugin marketplace flow.

Preview the generated curated install commands without installing:

npm run install:curated:dry-run

Preview the Codex skill sync without changing ~/.codex/skills:

npm run sync:codex-skills:dry-run

Install personal plus curated skills:

npm run install:all

Update globally installed skills managed by npx skills:

npm run update:skills

Curated Tools And References

Not everything belongs in skills/. Use curated-tools.json for tools, packages, apps, CLIs, and reference docs that agents should know I use, but that are not themselves agent skills.

Examples:

Tool	Type	How it is referenced
@google/design.md	npm package / CLI	Installed as a dependency; use npx @google/design.md lint DESIGN.md or npm run design:lint
rtk	CLI	rtk-ai/rtk; use as a token-saving proxy for verbose shell commands
codeburn	CLI	getagentseal/codeburn; use to visualize token usage, cost, models, and waste
gnhf	CLI plus skill	kunchenguid/gnhf; use for bounded long-running agent loops with concrete stop conditions
autoskills	CLI	Use npx autoskills or npm run autoskills for automatic context-aware skill discovery

@google/design.md is a good example: it is a design-system format and linter, not an agent skill. Keeping it in curated-tools.json lets this repo remember the tool and the commands without exposing it as something an agent should invoke as a skill.

Design.md commands:

npm run design:lint
npx @google/design.md diff DESIGN.md DESIGN-v2.md

RTK commands:

npm run rtk:verify
npm run rtk:stats

CodeBurn commands:

npm run codeburn:status
npm run codeburn:optimize

GNHF commands:

npm run gnhf:verify
gnhf --worktree --max-iterations 5 "improve this branch and stop when relevant checks pass"

Autoskills commands:

npx autoskills
npm run autoskills

AGENTS.md

AGENTS.md is the public global agent-instruction source of truth for this setup. It should describe durable cross-agent behavior, default skill routing, editing rules, and public-safe tool guidance.

It is not just repo-local guidance. Keep it suitable for reuse by global agent configuration flows and do not copy generated local memory context or private project instructions into it.

Reference vs Fork

Use curated-skills.json when the upstream skill works as-is.

Fork a third-party skill into skills/ only when I need to change its behavior. When forking, keep the original source in the skill body or references so the upstream lineage is clear.

Plugin Metadata

The .claude-plugin/, .cursor-plugin/, and .codex-plugin/ manifests describe this repo as a skills package for agents that understand plugin metadata.

The curated third-party skills are not bundled into those plugin manifests. They are installed separately through the package scripts so their upstreams remain clear and independently updateable.