Turn One Plan into Parallel Agent Work with CodexSkills

Some features are too big for one agent session: planning, backend, frontend, then verification. CodexSkills splits that work in two — a planner skill that writes phased implementation plans, and a runner skill that executes the plan file with parallel subagents.

Why This Skill Matters

CodexSkills (am-will/codex-skills, 1,033 stars in the latest leaderboard snapshot) is a collection of agent skills for planning, documentation access, frontend development, and browser automation. The README groups its skills into six categories: Agent Orchestration, Documentation Access, Prompting, Frontend & Design, Codex Tooling, and Browser Automation. The repo's skills/ directory held 22 skill folders as of September 14, including planner, plan-harder, parallel-task, and llm-council. One caveat from the same check: the README's list names a Context7 docs skill, ctx7old, that was not present in the directory that day, and the install example's --skill context7 flag matched no folder either — list the directory before installing that one.

Installation

Install with the skills CLI the README documents:

# List available skills before installing
npx skills add am-will/codex-skills --list

# Install the planning pair to user scope
# (the README's example names context7; this installs the verified pair instead)
npx skills add am-will/codex-skills --skill planner --skill parallel-task -g

Use -a claude-code -a codex to target specific agents, and -y to skip confirmation prompts.

Real Workflow: Add OAuth Login Without One Long Session

1. Plan first. Ask the agent to plan the feature: planner creates comprehensive, phased implementation plans with sprints and atomic tasks. Use plan-harder when you want the more detailed analysis variant.
2. Fan out. parallel-task executes a plan file by launching multiple parallel subagents to complete tasks simultaneously. It requires an existing plan file from planner — the two skills are designed as a pair.
3. Pull docs while subagents run. read-github converts github.com/owner/repo URLs to gitmcp.io/owner/repo for LLM-friendly access to repo docs. markdown-url prefixes any URL with https://markdown.new/ for a clean, Markdown-friendly view.
4. Verify in a browser. agent-browser is a Rust-based headless browser automation CLI from Vercel Labs with a snapshot/act pattern for agents. The README notes you don't need both browser tools and recommends agent-browser for speed and simplicity.

Plan adding OAuth login to this web app with planner, then execute the
plan file with parallel-task.

For genuinely contested design decisions, llm-council spawns multiple AI planners (Claude, Codex, Gemini) that generate independent plans, then a judge agent synthesizes the best approach. It includes a real-time web UI that auto-launches during planning sessions, and it requires API keys or active subscriptions for Claude/Anthropic, OpenAI, and Google — run ./setup.sh in the skill directory to configure.

The Hook Catalog

The hooks/ directory adds a separate layer: 51 ready-to-install Codex hook bundles in 11 categories, from automation (18 bundles) and security (4) down to testing (1). The README documents one scripted install path (a manual copy alternative is documented alongside it):

python3 hooks/aitmpl-codex/install-bundle.py \
  hooks/aitmpl-codex/<category>/<bundle> <target-repo>

There is also an agents/ directory of TOML-based multi-agent definitions — specialized roles like architect, frontend, backend, and security, plus worker tiers for parallel task execution.

Tips

• Run --list before installing to see what the directory actually holds — the README list and the folder set have drifted (see ctx7old above).
• gemini-computer-use requires a GEMINI_API_KEY; the README suggests asking Codex to help you set it up.
• parallel-task needs a plan file first — start with planner, not the runner.
• The hook bundles are adapted for the Codex event model; check hooks/README.md in the repo for compatibility details.

When Not to Use This

The skills target Codex-style agent workflows, and the hook catalog is written for Codex hooks specifically — if your agent has no Codex hook support, the hook catalog does not apply. And llm-council multiplies API spend across three providers; the plain planner + parallel-task pair is the cheaper default.

See the leaderboard for more skills.