Skip to content
A/Agent Skills
← Back to catalog

Selection guide / Utilities

Know the job.
Choose the tool.

Use one lead skill for the main task. Add a specialist only when it owns a separate phase.

Utilities and Agent Tools

Discover, create, and install

  • $pick-skills: route a request to the right installed skill or skills when the choice is unclear or the task spans several phases. It returns a plan and stops; it does not perform the work, and it routes to $find-skills itself when nothing installed owns a phase. User-invoked only, so type it: the agent will never offer it.
  • $find-skills: search for a capability that is not installed yet.
  • $writing-skills: write or sharpen a skill of your own — naming, trigger descriptions, and what belongs in the body. It is about the craft, so it works for any agent; $skill-creator scaffolds the Codex file layout instead.
  • $using-superpowers: the convention its family expects — invoke a matching skill before answering. Read it when a Superpowers skill behaves in a way the guides do not explain; it is a protocol, not a task skill, and it never competes with /pick-skills for routing a request.
  • $skill-creator (Codex only): create or improve a Codex skill.
  • $skill-installer (Codex only): install a Codex skill from a curated source or repository.
  • $plugin-creator (Codex only): scaffold or update a Codex plugin. Usually selected only when plugin work is explicit.

System helpers

  • $openai-docs (Codex only): current, cited guidance for OpenAI products and APIs.
  • $screenshot: capture the desktop, an application window, or a pixel region. Pick it for native apps and whole-screen captures; type /playwright-cli instead when the target is a web page, since a browser tool can size the viewport and wait for the page.
  • $review-agent (Codex only): a system-managed read-only reviewer normally delegated by another workflow.
  • $full-output-enforcement: request exhaustive output when completeness matters more than brevity.
  • $agents-are-coming (Claude Code only): a dummy skill that hypes up the arrival of agents. No task use; it exists to announce agents.
  • $wizard: generate an interactive bash wizard that walks a human through a manual procedure — third-party setup, a one-off migration, an A→B transition — capturing values into .env files and Actions secrets. Pick it when steps must happen outside the agent, on someone else's machine or account; not when the agent can simply perform the task itself.

Agent runtime and orchestration (Claude Code only)

Bundled by plugins, and mostly internal contracts rather than skills you invoke by hand — read them to understand how a plugin routes work, not to route it yourself.

  • $orchestration (fable-advisor plugin): the architect-as-orchestrator doctrine — when an Opus session delegates routine implementation to a cheaper lane and when it escalates. A decision protocol, not a task skill; it does not compete with /pick-skills for routing a single request.
  • $codex-cli-runtime (codex plugin): the internal contract for calling the codex-companion runtime from Claude Code.
  • $codex-result-handling (codex plugin): internal guidance for presenting Codex helper output back to the user.
  • $gpt-5-4-prompting (codex plugin): internal guidance for composing Codex / GPT-5.4 prompts inside the Codex plugin.
  • $shared (apple-skills plugin): the plugin's skill-creation templates and shared conventions. Pick it for authoring within that plugin's layout; $writing-skills above is the agent-neutral craft guide.
  • $knowledge-update (vercel plugin): corrects outdated model knowledge about the Vercel platform. Injected at session start, not invoked on demand.

Delegation and compressed communication

  • $cavecrew: decide when to delegate to compressed specialist subagents.
  • $dispatching-parallel-agents: run several independent subagents at once and merge their results. Pick it when the work genuinely splits — separate files, separate questions — and $cavecrew when the point is a compact answer from one specialist rather than fan-out.
  • $caveman: intentionally terse communication mode.
  • $caveman-review: extremely compressed code-review comments.
  • $caveman-commit: compact Conventional Commit messages.
  • $caveman-compress: compress a memory or instruction file while retaining a readable backup.
  • $caveman-stats: inspect actual token usage and estimated savings.
  • $caveman-help: display the Caveman command reference.
  • $ponytail-help: display the Ponytail command reference. For the mode itself and its review skills, see development.
  • $ponytail-gain: display Ponytail's published benchmark scoreboard. It is a fixed reference card, not a measurement of your repository, so do not pick it to size the impact of a change.

These utilities modify how work is routed or communicated; they generally should not replace the domain skill that owns the task.