agent-skills

Personal Codex and Claude Code agent skills for writing, plan review, Codex review loops, unattended merge-ready pull requests, pull request upkeep, browser-based GPT-6 Pro collaboration, and working with Claude Code CLI, OpenAI Codex CLI, Google Antigravity CLI, and Cursor CLI.

Available Skills

Skill	Description
antigravity-cli	General-purpose Antigravity CLI runner (agy, Google's successor to Gemini CLI) — code analysis, refactoring, and automated editing with model/autonomy/sandbox selection. Recommends Gemini 3.8 Flash when available, validates exact model slugs with agy models, and supports JSON output on current builds.
claude-cli	General-purpose Claude Code CLI runner (claude, claude -p) — code analysis, refactoring, reviews, automated editing, background agents, and structured output. Covers Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 4.5, with provider-specific alias guidance. Defaults to Opus; multiple Fable instances require explicit cost confirmation.
chatgpt-pro	Consults GPT-6 Pro when requested or proactively for stalled hard problems and consequential reasoning decisions. Works with Codex, Claude Code, and other agents with browser UI tools; preserves practical outputs and cross-agent handoffs. Prefers the user's default browser where ChatGPT is already logged in.
codex	General-purpose Codex CLI runner — code analysis, refactoring, and automated editing with model/effort/sandbox selection for GPT-6 Astra (gpt-6-astra), GPT-6.1 Sol (gpt-6.1-sol), and GPT-6 Luna (gpt-6-luna). Includes max effort and model-specific ultra support.
codex-loop	Reviews a change with Codex in a loop until both agents agree: up to 3 rounds by default, a findings ledger, and the smallest fix for each real finding. Runs through T3 Code delegation or Codex CLI.
codex-plan-review	Iterative plan review loop — Claude sends the current plan to Codex, revises based on feedback, and re-submits until Codex approves (up to 5 rounds), with GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna model options.
cursor-cli	General-purpose Cursor CLI runner (agent, cursor-agent) — code analysis, reviews, planning, refactoring, and automated editing with account-specific model discovery, ask/plan modes, permissions, structured output, and session resume. Suggests Composer, Muse, and Grok models; defaults to Composer.
leaving-pr-comments	Formats comments that an agent posts on GitHub on your behalf: an attribution callout, then a direct reply that leads with the conclusion.
merge-ready	Takes a change to a pull request with nobody watching: proves it, reviews it with codex-loop, files the PR, and babysits CI and review bots. Stops when the PR is ready for you to merge and never merges it.
modernize	Brings an old PR up to date and decides whether it should still exist: recommends modernize, rework, or close, then rebases and pushes only when the PR still belongs.
writing	Draft, rewrite, critique, or unblock human-facing prose while preserving facts, voice, and genre-specific intent.

Installation

Plugin version: 1.12.0. Model guidance checked on 2026-10-01; availability depends on the provider, account, and installed client. Explicit model selections are preserved.

skills.sh

Install all skills:

npx skills add makeavish/agent-skills

Install individual skills:

npx skills add makeavish/agent-skills --skill antigravity-cli
npx skills add makeavish/agent-skills --skill claude-cli
npx skills add makeavish/agent-skills --skill chatgpt-pro
npx skills add makeavish/agent-skills --skill codex
npx skills add makeavish/agent-skills --skill codex-loop
npx skills add makeavish/agent-skills --skill codex-plan-review
npx skills add makeavish/agent-skills --skill cursor-cli
npx skills add makeavish/agent-skills --skill leaving-pr-comments
npx skills add makeavish/agent-skills --skill merge-ready
npx skills add makeavish/agent-skills --skill modernize
npx skills add makeavish/agent-skills --skill writing

Some skills call others. When installing individually, install the ones they need too: merge-ready uses codex-loop and leaving-pr-comments, modernize uses leaving-pr-comments, and codex-loop uses codex outside T3 Code.

Claude Code

/plugin marketplace add makeavish/agent-skills
/plugin install personal@makeavish-skills

To update after new releases:

/plugin marketplace update
/plugin update personal@makeavish-skills

Codex

Add this repository as a Codex plugin marketplace, then install the plugin:

codex plugin marketplace add makeavish/agent-skills
codex plugin add personal@makeavish-skills

ChatGPT Pro requirements and continuity

The concise, agent-neutral chatgpt-pro skill requires browser UI tools, an authenticated ChatGPT session with GPT-6 Pro, and an authorized writable artifact location. Codex, Claude Code, and other hosts use their own available computer-use or browser tools. A host without those capabilities can prepare a blocked handoff but cannot consult Pro. Install or load the skill in each host that should use it; automatic discovery depends on that host's skill support.

For difficult work, the skill calls for a focused Pro consultation after local investigation identifies a concrete reasoning gap. Routine tasks and user requests to avoid external consultation stay local. Sessions normally save results and restart context under .agents/pro-sessions/<session-id>/HANDOFF.md. For a read-only repository, use an authorized durable directory outside it; if none is available, ask for one before browser actions or writes. A handoff can target any agent with access to the workspace and artifacts.

IDs

• Claude Marketplace ID: makeavish-skills
• Codex Marketplace ID: makeavish-skills
• Plugin ID: personal
• Repository: makeavish/agent-skills

Repository Structure

.
├── .agents/
│   └── plugins/
│       └── marketplace.json
├── .claude-plugin/
│   └── marketplace.json
├── plugins/
│   └── personal/
│       ├── .claude-plugin/
│       │   └── plugin.json
│       ├── .codex-plugin/
│       │   └── plugin.json
│       └── skills/
│           ├── antigravity-cli/
│           │   └── SKILL.md
│           ├── claude-cli/
│           │   └── SKILL.md
│           ├── chatgpt-pro/
│           │   ├── assets/
│           │   │   └── session-handoff-template.md
│           │   ├── evals/
│           │   │   ├── files/
│           │   │   │   ├── api-design-HANDOFF.md
│           │   │   │   ├── auth-review-HANDOFF.md
│           │   │   │   ├── auth-review-browser-mock.md
│           │   │   │   ├── incident-context.md
│           │   │   │   └── rollout.md
│           │   │   ├── evals.json
│           │   │   └── trigger-evals.json
│           │   └── SKILL.md
│           ├── codex/
│           │   └── SKILL.md
│           ├── codex-loop/
│           │   └── SKILL.md
│           ├── codex-plan-review/
│           │   └── SKILL.md
│           ├── cursor-cli/
│           │   └── SKILL.md
│           ├── leaving-pr-comments/
│           │   └── SKILL.md
│           ├── merge-ready/
│           │   └── SKILL.md
│           ├── modernize/
│           │   └── SKILL.md
│           └── writing/
│               ├── evals/
│               │   └── evals.json
│               └── SKILL.md
├── LICENSE
└── README.md

Adding a New Skill

1. Create plugins/personal/skills/<skill-name>/SKILL.md
2. Add frontmatter with name, description (controls auto-trigger), and skill instructions
3. Bump the version in both plugins/personal/.claude-plugin/plugin.json and plugins/personal/.codex-plugin/plugin.json
4. Update the skill table and any affected installation or structure documentation in this README
5. Run the Claude and Codex plugin validators
6. Reload or reinstall the plugin in the target client

plugins/personal/skills/my-skill/
└── SKILL.md

Conventions:

• name in frontmatter must exactly match the directory name
• description should explain both what the skill does and when it should trigger
• Move reference material into references/ subdirectory when the SKILL.md gets large

Credits

The codex-loop, leaving-pr-comments, merge-ready, and modernize skills come from t3dotgg/tstack by Theo Browne, used under the MIT license. merge-ready is adapted from full-stop and the skills it builds on, and codex-loop from sol-loop; the others are adopted with small changes.

License

MIT. See LICENSE.