v0.1.0 — prompt optimizer by official vendor rules

Rewrite your prompt into the native dialect of the model you're actually running — Claude Code or Codex — using that vendor's official rules, and see which rule drove each edit.

The same prompt gets opposite treatment: think step by step is kept for Claude and stripped for Codex reasoning models; double-check yourself is removed on Opus 5 (it self-verifies) but left for Codex. Every suggestion links to the vendor doc it came from.

What's in 0.1.0

• improve — rewrite + explain, each edit linked to an official rule
• detect — which model you're on, and where that answer came from (settings cascade, VS Code aware, [1m] context suffix preserved, aliases never guessed into a version)
• rules — every rule with its source, plus the model coverage table
• update --diff — watches 14 official vendor pages and shows the exact before/after lines; weekly CI opens a PR with that diff
• Run-mode advice: when /goal, plan mode or a dynamic workflow — and an honest "Codex has no /loop"
• Claude Code skill, UserPromptSubmit hook, and a Codex skill/plugin

Coverage — Claude: 11 model generations (4 have their own vendor prompting page; the rest run on family rules, not invented ones). Codex: GPT-5.6 family (sol/terra/luna share one official guide).

Zero runtime dependencies, stdlib only, works offline, no API key. 67 tests. See CLAIMS.md for what it does not do.