agent-instructions

Skills

This repository holds my reusable skills and workflows for AI agents, packaged as Claude Code and Codex plugins. Every plugin is also an Agent Plugins 1.0.0 package, so any client that reads that standard can load it. Use them directly, or explore them for inspiration.

Install

You can install plugins from this repository in both Codex and Claude Code. Every plugin is available to both agents unless its section below says otherwise.

Codex

Add the marketplace to Codex:

codex plugin marketplace add 0xA3B/skills

Upgrade the marketplace later:

codex plugin marketplace upgrade 0xa3b-marketplace

After you add or upgrade the marketplace, install the plugins that you want from Codex.

Claude Code

Add the marketplace inside Claude Code:

/plugin marketplace add 0xA3B/skills

Then install the plugins that you want:

/plugin install codex-in-claude@0xa3b-marketplace
/plugin install git@0xa3b-marketplace
/plugin install engineering@0xa3b-marketplace
/plugin install writing@0xa3b-marketplace
/plugin install meta@0xa3b-marketplace

Update the marketplace later with /plugin marketplace update 0xa3b-marketplace.

Plugins

claude-in-codex

Codex-only: this plugin exists to drive Claude Code from Codex, so this repository does not publish it to the Claude Code marketplace.

This plugin runs Claude Code from Codex for adversarial code review, feedback triage, and in-scope fixes.

• claude-in-codex:adversarial-review: Invokes Claude Code as a review-scoped adversarial reviewer, triages findings as external feedback, and applies accepted in-scope fixes for current working-tree changes.
• claude-in-codex:using-claude-cli: Internal contract for running the Claude Code CLI non-interactively, with prompting guidance and a copyable .codex/agents/claude.toml proxy subagent definition.

codex-in-claude

Claude Code-only: this plugin exists to drive Codex from Claude Code, so this repository does not publish it to the Codex marketplace.

This plugin runs Codex from Claude Code for adversarial code review, feedback triage, and in-scope fixes. It also ships a codex subagent that delegates tasks to Codex.

• codex-in-claude:adversarial-review: Invokes Codex as an adversarial reviewer at the user's configured sandbox and permission defaults, triages findings as external feedback, and applies accepted in-scope fixes for current working-tree changes.
• codex-in-claude:using-codex-cli: Internal contract for running the Codex CLI non-interactively, with GPT-5.5 prompting guidance.

git

These skills create Conventional Commits and drive GitHub or GitLab change requests from operational preparation through automated review and verified merge cleanup.

address-pr-feedback requires the engineering plugin so it can apply the shared receiving-feedback discipline.

• git:commit: Reviews current changes, stages logical units, and creates git commits with Conventional Commit messages.
• git:tidy-history: Folds, splits, reorders, and rewords a topic branch's commits without changing its final tree, then force-pushes the result with --force-with-lease without asking; on main, the remote's default branch, or the change request's target it only reports its plan.
• git:create-pr: Prepares a branch, creates or refreshes its pull request or merge request, and observes initial CI.
• git:address-pr-feedback: Drives active automated-review adapters to approval or a reported exception.
• git:merge-pr: Verifies merge gates, merges synchronously, and cleans up verified local and remote branch state.

engineering

These skills cover the engineering loop from decision mapping and design review through implementation, diagnosis, code review, and session handoff. Some skills come from Matt Pocock's MIT-licensed mattpocock/skills repository; each adapted skill records its original author and source in its Agent Skills frontmatter metadata.

The typical implementation flow starts with wayfinder, which maps a loose idea into decision-sized chunks. grill-me resolves a selected direction, and prototype answers questions that need executable evidence and loads on its own for prototype and spike requests. tdd implements observable behavior through a red-green-refactor loop over a frontier of behaviors, and loads on its own for implementation requests. review-changes reviews changes the session authored through focused lanes, fixes accepted findings, and reruns the lanes those fixes invalidated; tdd applies it at completion. codebase-design provides shared background discipline, and decision-records both backs the other workflows and loads on its own for requests to record or check a decision. Use handoff when another agent session should continue from an ignored local context document.

• engineering:codebase-design: Applies shared deep-module and interface-design discipline.
• engineering:decision-records: Records, updates, or supersedes architecture decision records, and reads the records that touch an area before a change there.
• engineering:dependency-maintenance: Reviews dependency update PRs, merges ready ones, syncs local state, refreshes repo-pinned tooling, and files linked follow-up issues.
• engineering:diagnose: Runs a disciplined diagnosis loop for bugs, flaky behavior, and performance regressions.
• engineering:grill-me: Stress-tests a plan, decision, idea, or design through adversarial questioning.
• engineering:handoff: Saves a compact local continuation document for another agent session.
• engineering:improve-codebase-architecture: Finds focused, evidence-backed module deepening opportunities.
• engineering:improve-codebase-tests: Finds evidence-backed test-suite improvement opportunities.
• engineering:prototype: Creates disposable exploratory code to answer a design question.
• engineering:review-changes: Reviews changes authored in the current session through focused lanes, fixes accepted findings, and reruns invalidated lanes.
• engineering:receiving-feedback: Triages and responds to existing review feedback.
• engineering:tdd: Builds features or fixes with a red-green-refactor loop.
• engineering:terminology: Creates, updates, or reviews durable project terminology.
• engineering:wayfinder: Maps a loose idea into decisions, dependencies, frontier, fog, and excluded scope.

meta

This plugin maintains the marketplace itself; its skills operate on the marketplace's own skills and repository rather than on a user's project.

• meta:submit-skill-feedback: Captures the session's qualified feedback from an actual marketplace-skill run. It files public-source feedback a