The repository curates Codex agent prompts rather than executable source code. Author reusable prompt blueprints under codex/prompts/ as standalone Markdown files, and keep the aggregated registry in codex/prompts.json. The claude/ directory stays available for future agent profiles; leave it untouched unless scoped work requests otherwise. Organize new prompts by functional focus and ensure each prompt file name matches the slash command you expect to trigger inside Codex.
No compilation is required, but validate edits quickly:
# Confirm catalog structure and spot missing prompts
Get-ChildItem codex/prompts
# Lint JSON structure before committing
jq empty codex/prompts.json
# Preview prompt contents
Get-Content codex/prompts/<prompt-name>.mdUse Windows PowerShell modules already available; avoid introducing project-specific toolchains unless discussed.
Write prompts in polished professional English, targeting senior LLM practitioners. Follow the Codex custom prompt spec: use $1..$9 placeholders only when arguments are required, and escape literal dollar signs with $$. Maintain 80–100 character line wrapping for readability, heading levels that mirror task hierarchy, and action-oriented bullet lists. Keep JSON metadata synchronized: category names, agent names, and descriptions remain in Simplified Chinese, while prompt entries reference the corresponding English Markdown filename.
Manually exercise new prompts in a fresh Codex session to ensure placeholders render correctly and instructions produce the intended workflow. Maintain parity between prompts.json metadata and the Markdown body—if a prompt introduces parameters or dependencies, explicitly document them in both places. When updating multiple prompts, review diffs with git diff to confirm only the intended sections changed.
Adopt the concise style seen in recent history, e.g., docs(codex): 更新 prompts catalog. Summaries should describe intent (“完善性能诊断提示词”), with optional scope tags in parentheses. Reference related issues when available and attach representative screenshots of Codex runs if behavior changed. Pull requests should enumerate new or revised prompts, call out testing performed, and highlight any guidance for downstream agents.
- Draft the Markdown prompt with context, goal, workflow, and guardrails tailored to this repository’s agents.
- Update
codex/prompts.json, aligning the Chinese metadata with the Markdown filename and ensuring categorization matches the prompt’s purpose. - Validate formatting (Markdown preview,
jqfor JSON), run a fresh Codex session to smoke test, then prepare a focused documentation commit.