SuperAPI AI Directory

codex-prompting-local

Coding & development

View repository

Updated 10-15 22:15

About this skill

codex-prompting-local

This skill adapts the portable parts of codex-plugin-cc to the current nix/home-manager/agents/ architecture.

Use it for prompt and output contracts, not as a new runtime control plane. This repo's runtime authority still lives in Nix-managed hooks, agents, MCP definitions, and generated home files.

Core rules:

  • Use one concrete task per Codex run unless there is a clear reason to batch.
  • Prefer compact XML-style prompt blocks over long freeform instructions.
  • Tell Codex what done looks like, how to verify it, and when to keep going by default.
  • Use machine-readable review output only when the downstream consumer benefits from it. Otherwise keep the repo's normal findings-first prose review style.
  • Keep resumable results thread-aware and action-oriented. If a thread ID is known, surface it explicitly. If not, do not invent one.
  • Treat old slash-command behavior as workflow intent to preserve in skills, scripts, or prompts. Do not reintroduce plugin frontmatter, AskUserQuestion, or .claude-plugin assumptions.

Recommended workflow:

  1. Pick the smallest prompt shape that fits the task.
  2. Add only the blocks the task actually needs.
  3. Add a structured review contract only for review surfaces that should emit parseable output.
  4. For long-running or multi-turn work, write the result so it can be resumed cleanly from saved handoff context.

Open these references as needed: