Skills

Agent skills that add meaningful value beyond what an unaided model does out of the box, for work that matters to me.

A frontier model is very competent at most things, but its default in any domain is the competent average version of the thing. Each skill here is an effort to drag the model off that median in a chosen direction: narrow, stack-specific disciplines with explicit boundaries and proof obligations.

They carry my taste and intent, stack-specific decisions, failures observed in real work, and the evidence I require before calling work done. The agent gets more specific instructions for the work I care about without putting all of that back into every prompt.

How it fits together

+---------------------------------------------+
| Self-contained skills                       |
| judgment / boundaries / proof obligations   |
+----------------------+----------------------+
                       |
       +---------------+---------------+
       v               v               v
  Claude Code        Cursor          Codex
       +---------------+---------------+
                       |
                       v
            Work in target repository
                       |
                       v
               Project evidence
     source / packages / commands / rendered behavior

The skill supplies the discipline; the target repository supplies the facts. A plugin manifest per agent exposes the same skill source to each of them, while hooks, fixtures, CI, and evaluation tools make broken or stale guidance visible.

Each skill is self-contained. Install the repository once, then let the agent select a relevant skill from the task or name the skill directly.

Use a skill

After installation, start a new agent session so its skill catalog is rebuilt. Then ask normally:

Use improve-layout to repair the mobile overflow without changing the design.
Run openreview across this whole codebase, then plan the selected findings.
Use openreview to implement this named Next.js cache change and re-audit it.
Run quality-audit against this repository and report only verified findings.
Run quality-audit with the launch profile and get this deployment ready.
Run web-launch-checklist for this production release and produce a go/no-go record.
Use ship to commit this change and stabilize the pull request, but do not merge it.

The catalog below is grouped by who decides when a skill runs. Ambient skills apply whenever the work is in their domain, without being asked for. Actions activate from their descriptions when the task matches, or when you name one. Commands only ever run when you invoke them because their cost, timing, or blast radius needs the owner's decision.

The skills inspect and follow the target project's own source, installed packages, components, tokens, and verification commands. Installing this repository does not add a UI runtime, Tailwind configuration, or application dependency to the projects where the skills are used.

Skills

The openreview skill ships the review, planning, implementation, and re-audit workflow, not the openreview-next executable. It runs one initial scan and compacts the JSON report before model review; “whole codebase” changes scan scope, not manual review depth. Full deterministic coverage currently requires a separately provisioned CLI implementing OpenReview DiagnosticReport v1 and its v1 changed-scope comparison envelope. A trusted OpenReview source checkout outside the target may supply its absolute CLI entrypoint and declared runtime when its exact revision is recorded and rules list proves the command surface; the bounded build:scanner CI artifact is not that interface. Until the CLI has a distributable release, installations without either provisioned form must report scanner coverage as Unverified and may perform only clearly labelled advisor work; they must not run a target repository's package script as a substitute.

Install

This repository ships compatible skill metadata for Claude Code, Cursor, and Codex.

Agent Skills CLI

Use the portable installer for Cursor or another supported agent-skills harness:

npx skills add claudialnathan/skills

Codex plugin

Add this repository as an independent Git marketplace, then install the plugin:

codex plugin marketplace add claudialnathan/skills
codex plugin add skills@claudia-skills

Pull later revisions in two steps: refresh the Git marketplace snapshot, then rewrite the installed plugin cache from that snapshot:

codex plugin marketplace upgrade claudia-skills
codex plugin add skills@claudia-skills

Claude Code plugin

This repository is its own claudia marketplace. .claude-plugin/marketplace.json publishes the repository root as a versionless, commit-SHA plugin, so adding the repository is the whole setup:

claude plugin marketplace add claudialnathan/skills
claude plugin install skills@claudia

The same two steps work as /plugin marketplace add and /plugin install inside a session.

Pull later revisions with:

claude plugin marketplace update claudia
claude plugin update skills@claudia

A claudia marketplace registered from anywhere else has to go first, because a name can only be registered once:

claude plugin marketplace remove claudia

Git marketplaces are a Claude Code surface. The Claude apps and claude.ai install plugins through Organization settings > Plugins, which syncs a marketplace repository rather than reading one per user; a plugin distributed that way must keep executables out of a top-level bin/ directory, which this repository does.

Restart the relevant agent after installing or updating. Existing sessions keep the catalog they started with.

Repository systems

This is the source map for the skill library and the machinery around it. Only the skills/ row ships as skill content; nothing else installs itself into a target repository. Use Node 22 for the executable tooling.

License

MIT