better — Prompt Forge

better

Prompt engineering, in one keystroke.

A Claude Code skill that forges a half-formed idea into a precise, self-verifying prompt — instantly. No five-round interview. One line in, a prompt the model can execute on the first try and drive to done on its own.

When you work with a capable model, the prompt is the interface. A vague prompt makes it guess — wrong stack, wrong scope, happy-path only — and you pay for the guess in rework. A well-engineered prompt makes the right answer obvious and lets the run check its own work.

Prompt engineering is the skill that closes that gap. better makes it a reflex: you type /better <whatever you've got>, and it hands back a prompt built the way Anthropic's own documentation says to build one — specific, structured, and carrying a verifiable success criterion so the run loops to done instead of stopping at "looks plausible."

The whole point is speed: you reached for a slash command, so it forges now. It infers the gaps, states its assumptions out loud, and never interrogates you. Rough idea → strong prompt, in one shot.

_{Contents · The difference · Install · Usage · Anatomy · Principles · Provenance · Contributing}

The difference

Same eight words, typed by a developer in a hurry. On the left, what the model actually receives without better. On the right, what /better forges from it — instantly.

You type: /better add password reset to my app

add password reset to my app

Add a password-reset flow to <app>. Read @auth/
and @package.json first and match the existing
session + mailer patterns — don't add a new lib
if one's already wired.

Flow: request reset by email → single-use token
(hashed at rest, 1h expiry) → reset form checks
the token → password updated, all existing
sessions invalidated, confirmation email sent.

Constraints: identical response whether or not
the email exists (no account enumeration);
rate-limit per email and per IP; tokens are
one-time and expire. Don't touch the login flow.

Approach: read the existing auth code, sketch
the token model, then build test-first.

Success: one test drives the whole path —
request → token consumed → password changed →
old sessions rejected → token can't be reused;
invalid / expired / used tokens all covered.
Suite + typecheck green before you call it done.

Look at the last paragraph on the right. That success criterion is the engine. It turns an open-ended request into a goal the run can verify against — so it self-corrects to done instead of handing you something that merely compiles. Everything else better does is in service of producing that line, fast.

See the gallery — build, fix, refactor, write, research, data, image — in examples/before-after.md.

Install

Two ways. Pick whichever matches how you run Claude Code.

As a plugin — recommended

/plugin marketplace add realdeal88/better
/plugin install better@better

/better is live immediately. Update later with /plugin marketplace update better.

As a manual skill — no plugin system

git clone https://github.com/realdeal88/better.git
mkdir -p ~/.claude/skills
cp -R better/skills/better ~/.claude/skills/better