GitHub: carlkibler/agent-skills
A collection of skills for Claude Code and Codex that I've built and use daily. Skills are prompt-defined behaviors that AI agents invoke automatically based on context — a structured way to encode workflow knowledge so you don't explain the same thing in every session.
This repo is installable as a Claude Code plugin marketplace, a Codex plugin set, or just cloned and pointed at directly.
What's Here
Better Products
Skills for finding problems before your users do:
- pre-mortem — Multi-agent ranked risk analysis with mitigations. Run this before shipping anything important.
- first-run-red-team — Red-teams the onboarding and first-run experience specifically. Catches abandonment traps that normal review misses.
- support-inbox-simulation — Simulates the support emails and refund requests a launch will generate.
- trust-audit — Audits permissions, privacy behaviors, billing flows, and silent failures for things users will find creepy or hostile.
- empathy-audit — Four-lens review: user, machine, developer, support.
Dev Workflow
- handle-pr — Autonomously reads PR review comments, implements the changes, and reports back. Works across multiple comment threads.
- second-opinions — Routes a question or diff to a second AI model and surfaces the disagreements.
- parallel-isolated-app-testing — Designs parallel test lanes for apps with shared local state (databases, file systems, queued jobs).
Utilities
- chezmoi-drift — Audits chezmoi dotfiles for drift, broken symlinks, and misconfigured skills.
- profile-me — Builds a portable AI context profile from your digital footprint.
- wifi-qr — Generates a WiFi QR code PNG.
Install
/plugin marketplace add carlkibler/agent-skills
/plugin install pre-mortem@agent-skills
Or browse in the UI: /plugin → Discover tab. Run /reload-plugins after installing.
For Codex: the repo includes native Codex packaging under plugins/ and .agents/. Clone the repo, open it in Codex, type $ or /skills.
Why Skills
The problem they solve: I was writing the same context into every Claude session. "Before you ship anything, think about what could go wrong. Check the first-run experience. Look for things users will find creepy." That's useful, but it's expensive to re-establish and easy to forget.
Skills encode that knowledge once and invoke it automatically. The pre-mortem skill, for example, spawns multiple agent perspectives (pessimist, user advocate, ops engineer), runs them in parallel, and synthesizes a ranked risk list. Doing that manually in a conversation is possible but awkward. As a skill it takes one invocation.
The repo also serves as a reference for how to write skills that work well across Claude Code and Codex — the packaging, the metadata, the distinction between local-machine skills and cloud-friendly ones.