wdym

# wdym

> A prompt rewriter for Claude Code and Codex that fires automatically on every prompt you submit, detects its type, and rewrites it with up to three matched principles — or leaves it alone when nothing would improve it. Flash-only: the rewrite runs at once, with no approval gate.

wdym is a skill plus a `UserPromptSubmit` hook, running on Claude Code (as a plugin or a skill) or Codex. The hook
(`prompt-detect.py`) scores each prompt deterministically and — for substantive prompts —
injects a host-neutral `<prompt-detect>` signal block; a trusted contract in the host's
instruction file (`CLAUDE.md` on Claude Code, `AGENTS.md` on Codex, written by the
installer or `/wdym --init`) tells the model to run the skill once per block before answering. The skill
classifies the prompt (`code` / `question` / `text-gen` / `none`), picks the
principles that clearly apply from `refs/protocol.md` § Principles, and rewrites —
never inventing facts, never adding placeholders, never changing the task. To change behavior, start at `SKILL.md` (a thin
dispatcher) and follow it into `refs/protocol.md`. Most detail lives in `refs/*` and loads
lazily — the SKILL.md body is re-injected on every run, so it stays small on purpose.

## Docs

- [README](https://raw.githubusercontent.com/ndisisnd/wdym/main/README.md): what it does, install, how it works, flags
- [SKILL.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/SKILL.md): the router — start here to modify behavior; dispatches into refs/
- [ARCHITECTURE.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/ARCHITECTURE.md): full flow diagram, file map, and design notes (token budget, one-file run, degrade-then-heal, two-host model)

## Reference

- [refs/protocol.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/refs/protocol.md): step-by-step execution, Steps 0–7, plus § Principles — the one file a run reads, and the place to add custom principles
- [refs/init.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/refs/init.md): --init bootstrap — the ask step (AskUserQuestion or plain text), host resolution, scope, activation, hook wiring, trust contract
- [refs/detect.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/refs/detect.md): manual type-detection path, used only when the hook verdict is ambiguous or absent
- [refs/commands.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/refs/commands.md): /wdym command handling (--init / --help / --status)
- [refs/categories.json](https://raw.githubusercontent.com/ndisisnd/wdym/main/refs/categories.json): the type taxonomy — keyword/regex scoring the hook and skill read (user-editable)
- [hooks/prompt-detect.py](https://raw.githubusercontent.com/ndisisnd/wdym/main/hooks/prompt-detect.py): the UserPromptSubmit pre-scorer that emits the <prompt-detect> block
- [.claude-plugin/plugin.json](https://raw.githubusercontent.com/ndisisnd/wdym/main/.claude-plugin/plugin.json): Claude Code plugin manifest; hooks/hooks.json wires the plugin hook
- [install.sh](https://raw.githubusercontent.com/ndisisnd/wdym/main/install.sh): the installer — host detection (Claude Code and/or Codex), skill files, hook wiring, trust contract, pref.json; `--uninstall` reverses it

## Optional

- [RELEASES.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/RELEASES.md): what changed for users, release by release
- [CHANGELOG.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/CHANGELOG.md): version history
- [LICENSE.md](https://raw.githubusercontent.com/ndisisnd/wdym/main/LICENSE.md): MIT