prompt-preflight

Rewrite rough, vague prompts into clearer, better-structured ones with Claude Haiku before they reach a stronger model — through a Claude Code proxy, a UserPromptSubmit hook, or a standalone CLI.

You type the way you always do. A small, fast model does a pre-flight pass that sharpens scope and structure without inventing requirements, and only the improved prompt is sent onward. If anything goes wrong, it fails open to your original text — the tool can slow you down by at most a second, never block you.

flowchart LR
    you([Your rough prompt]) --> pf[prompt-preflight<br/>Claude Haiku]
    pf -->|clearer prompt| strong([Strong model<br/>Opus / Sonnet])
    pf -. fail-open .-> strong

Why

A stronger model rewards a clearer prompt. But stopping to hand-craft every prompt is friction. prompt-preflight removes that friction: it spends ~1 second and a fraction of a cent on Haiku so your expensive model spends its budget on a well-framed request.

Three ways to use it

Path	What it does	Best for
Proxy (enhance)	True replacement — Claude Code is routed through a local proxy that rewrites the request body	Interactive Claude Code
Hook (enhance-hook)	Injects a clarified restatement as additionalContext	Zero-setup fallback
CLI (enhance-cli)	Rewrites text and copies it to your clipboard	Desktop / web apps

Core guarantees

• Faithful. A programmatic check verifies your hard tokens (paths, URLs, code spans, numbers) survive the rewrite; if any are dropped, it fails open.
• Private. Nothing is written to disk by default. Diagnostics are opt-in and local-only. Credentials are detected and never sent to the enhancer.
• Safe. Fail-open everywhere, a kill-switch (PROMPT_ENHANCER_DISABLE=1), and a //raw bypass to skip enhancement on demand.

Start with the Quickstart.