Agent-Skills.md

Agent Skills: Codex Utils

Discover available tools, coordinate safe parallel reads, clarify decisions, inspect images, and verify current sources.

UncategorizedID: bjornmelin/dev-skills/codex-utils

Repository

bjornmelin/dev-skills
BjornMelin
8

Install this agent skill to your local

pnpm dlx add-skill https://github.com/BjornMelin/dev-skills/tree/HEAD/skills/codex-utils

Skill Files

Browse the full folder contents for codex-utils.

Download Skill

Loading file tree…

skills/codex-utils/SKILL.md

Skill Metadata

Name
codex-utils
Description
Discover available tools, coordinate safe parallel reads, clarify decisions, inspect images, and verify current sources.

Codex Utils

Use the lightest available tool that resolves the need. The current session's tool catalog, schemas, and restrictions are authoritative; never assume a tool name from another host or session exists.

Coordination

  • Use a planning tool when available and useful for non-trivial work; otherwise keep a short checklist. Do not create a persistent goal unless requested.
  • Use the available user-input tool for optional clarification, following its current mode restrictions. Ask directly when explicit approval is required; a recommended or preselected option is not approval.
  • Inspect local images with the available image-viewing tool before editing or judging them. Use original detail only when necessary.

Discovery and execution

  • Search the exposed tool catalog before declaring a capability missing. In Code Mode, filter ALL_TOOLS by name/description and call the matching tools method. Use tool search only when the session exposes it.
  • Match the actual schema before calling. Never print credentials, raw auth configuration, or secret-bearing subprocess arguments while inspecting tools.
  • Batch independent read-only calls with Promise.allSettled when Code Mode is available; inspect every result. Keep dependent calls, mutations, approvals, waits, and tools marked non-parallel sequential.
  • Request plugin installation only for a plugin or connector explicitly requested by the user, after the available discovery/install tools confirm an exact match. Installed, enabled, exposed, and healthy are different states; verify the one the task needs.

Sources

Use the available browsing tool's search/open/find operations for current facts and source verification. Prefer primary sources; inspect the actual page before relying on search snippets. For OpenAI product questions, inspect local code/help first and use official OpenAI documentation as the fallback. Keep private source and credentials out of external research services.

Defaults

Verify unstable or high-stakes claims with tools instead of memory. Keep tool use proportional: do not browse, ask, plan, install, or delegate when unnecessary.