ai-coding-agent-skills

ai-coding-agent-skills

Agent skills for AI coders

Overview

Single-source, reusable skills and agent prompts shared across AI coding runtimes. The skills/ directory is the source of truth and is referenced by runtime-specific integrations via .claude/skills -> ../skills and per-skill symlinks under .agents/skills/ (e.g. .agents/skills/<skill> -> ../../skills/<skill>). Autonomous or scheduled workflows should invoke these canonical skills instead of maintaining duplicate routine prompts.

Each skill directory contains a SKILL.md that documents prerequisites and invocation.

Quick start

1. Clone the repo:

   git clone git@github.com:dceoy/ai-coding-agent-skills.git
   

2. Pick a runtime and explore the relevant integration:

   • Claude Code / Claude Code Routines: .claude/skills -> ../skills (symlink exposing all skills)
   • Codex CLI skills: .agents/skills/ (per-skill symlinks into skills/)
   • Codex CLI custom subagents: .codex/agents/ (project-scoped definitions discovered automatically by current Codex releases)

3. Open a skill directory or .codex/agents/README.md to learn how to invoke it.

Skills

All skills are located in skills/ and surfaced through shared discovery or runtime-specific symlinks.

Git Workflows

• clean-gone-branches - Clean up local branches marked as [gone] and their worktrees
• commit - Create a git commit with an appropriate message
• commit-push-pr - Commit, push, and open a pull request

Code Quality

• parameterized-tests - Prefer native parameterized or table-driven tests for repeated unit-test cases that share the same test logic
• simplify-codebase - Reduce maintenance surface under KISS/DRY/YAGNI, explicitly or proactively when worthwhile

AI Tools

• oracle-chatgpt - Send one arbitrary prompt to ChatGPT through Oracle browser mode and return the captured response without reinterpretation
• x-timeline - Read authenticated X timelines through agent-browser without engagement actions

Standalone Tools

• scripts/oracle-pr-sweep.sh [--debug] [--max-count=<int>] - Submit an open-PR batch review through Oracle + ChatGPT @GitHub, requesting one COMMENT review per current PR. Runs Oracle in the foreground in browser mode with gpt-6 (requires Oracle 0.20.1+ and GPT-6 browser access), waiting for the browser session to finish. The CLI exit status does not independently confirm that GitHub reviews were published. The wrapper retains no separate status or output log; use oracle status or oracle session <id> for Oracle-managed runs. The positional [MAX_COUNT] form remains supported; the default limit is 20 (range 1-50).

Delivery Analytics

• github-productivity - Collect, retain, normalize, and analyze GitHub organization PR/review/timeline/commit data via gh api into an organization-week delivery panel with a pre-specified interrupted time-series model, sensitivities, charts, and a report

Artifact Deployment

• gas-html-artifact - Deploy an existing HTML artifact to Apps Script HTML Service with clasp, preserving compatible bytes and existing project files
• gas-drive-html-server - Deploy an independent Drive-backed HTML directory index as a GAS Web App, with live folder listings and no HTML redeploys

Skill Management

• claude-agent-converter - Convert Claude Code agents to portable skills
• claude-command-converter - Convert Claude Code commands to portable skills

Related projects

For issue-to-PR implementation, pull request review, and feedback resolution, see dceoy/pr-loop. It is maintained separately and provides the issue-to-pr, pr-review, pr-feedback-triage, and composite pr-loop agent skills.

Codex Custom Subagents

Four optional native, read-only roles are defined under .codex/agents/: planner, advisor, reviewer, and feedback-analyst. Each TOML specifies its default model and reasoning effort; the top-level agent owns implementation. The routing policy is .codex/AGENTS.md, which must be installed as $CODEX_HOME/AGENTS.md to apply globally (it is not read automatically from the repository root).

See Codex agent setup for installation and capability checks. Unsupported named-role dispatch falls back to work in the main agent, not a simulated child agent.

Structure

.
├── skills/                  # Shared skill directories (source of truth)
├── .agents/
│   └── skills/              # Per-skill symlinks into skills/ (other runtimes)
├── .claude/
│   └── skills -> ../skills  # Symlink exposing skills/ to Claude Code runtime
├── .codex/
│   ├── AGENTS.md            # Codex user-wide routing template
│   └── agents/              # Project-scoped Codex custom subagents
├── .github/
│   └── workflows/           # CI workflows (ci.yml)
├── scripts/                 # Standalone CLI helpers
├── README.md
└── LICENSE

Prerequisites

Install and authenticate the required CLI tools before running skills:

• Node.js >=20 and Google's clasp CLI - For gas-html-artifact and gas-drive-html-server

  • Use pnpm exec clasp (preferred), npx --no-install clasp, or direct clasp depending on available executables
  • Authenticate with the chosen launcher and enable the Apps Script API before deployment
  • Mocked deployment tests require Node.js but no Google credentials

• Claude Code - For .claude/ agents and skills

  • Install: https://docs.anthropic.com/en/docs/claude-code
  • Auth: Follow CLI onboarding flow

• Codex CLI - For .agents/skills/ and .codex/agents/

  • Install: npm install -g @openai/codex
  • Auth: run codex login

• Oracle CLI - For oracle-chatgpt and scripts/oracle-pr-sweep.sh

  • Install: npm install -g @steipete/oracle
  • ChatGPT: sign in for Oracle browser mode
  • Remote browser routing is optional and uses Oracle's native configuration

• agent-browser - For x-timeline

  • Install: npm install -g agent-browser (or the package manager your environment uses), so agent-browser is on PATH
  • No custom wrapper, browser service, or MCP server is required or bundled; x-timeline invokes agent-browser directly and fails closed if the installed workflow cannot enforce its documented safeguards.
  • x-timeline prefers a reusable dedicated X session already on https://x.com/home, so routine reads need no navigation or tab-selection action. Initial setup and tab changes use guarded actions and capability-check the installed agent-browser workflow instead of pinning behavior to a specific upstream release number.

Usage notes

• Skills do not always auto-run; use your agent's skill invocation flow or ask for the skill explicitly.
• For Claude Code Routines, CI, and other autonomous workflows, invoke the canonical skill under skills/ and pass runtime-specific context externally.
• If a skill fails, open its SKILL.md and verify prerequisites and command syntax.

Troubleshooting

Skill not found

• Confirm the skill directory exists in the expected runtime location
• Check that skill name matches exactly (case-sensitive)
• Verify the SKILL.md documentation is present

CLI not in PATH

• Ensure the tool is installed and accessible: which <tool-name>
• Add the tool's bin directory to your shell PATH
• Restart your terminal after installation

Authentication errors

• Re-run the tool's auth command:
  • Claude Code: Follow onboarding flow
  • Codex CLI: run codex login
• Verify active subscription or API key

Symlink issues

• Skill directories are shared from skills/ via .agents/skills and .claude/skills
• If broken, recreate the symlink or ensure skills/ exists

License

See LICENSE for details.