Turn a rough, half-formed idea into a powerful, execution-ready prompt — fast. Use whenever the user types /better, or says "make this prompt better", "improve my prompt", "rewrite this so the model nails it", "turn this into a good prompt", "prompt-ify this", or hands over a vague one-liner they want sharpened before running it. Also use proactively when a request is clearly an under-specified prompt the user intends to feed to an AI and would get a far better result if it were forged first. Forges prompts grounded in Anthropic's own prompting guidance and tuned for a high-effort reasoning model (Opus 4.x): every forged prompt carries a verifiable success criterion so the downstream run drives itself to done without babysitting. Optimized for one-shot slash use — forge instantly, ask only when a gap would genuinely change the output.
How this skill is triggered — by the user, by Claude, or both
Slash command
/better:betterThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Turn a messy idea into a prompt so clear and well-scoped that the downstream run is effective on the first try and can drive itself to "done" without babysitting. The user reached for a slash command — they want this *now*, not after a five-round interview.
Turn a messy idea into a prompt so clear and well-scoped that the downstream run is effective on the first try and can drive itself to "done" without babysitting. The user reached for a slash command — they want this now, not after a five-round interview.
This skill encodes how Anthropic actually teaches prompting: state success criteria instead of procedures, wire a way for the run to check its own work, be specific, pair every hard requirement with a mechanism, and give a high-reasoning model room to plan. Type-specific skeletons, the craft's reasoning, and a worked-example library live in references/forge-anatomy.md — read it whenever you forge something non-trivial or want a type-specific structure.
A weak prompt makes the model guess; a strong one makes the answer obvious. Most rough ideas already carry enough signal to forge — you fill the gaps with smart, stated assumptions instead of stopping to ask. Vagueness is the expensive path: a thin prompt forces the downstream run to burn context reconstructing what you meant ("just make it better" is the canonical bad prompt).
Default: forge immediately. Only pause when a missing detail would genuinely change what gets built and you can't pick a sane default. When you must ask, ask once via the AskUserQuestion tool (clickable options), 2–4 questions max, then forge. Never run multiple question rounds — that defeats the point of a slash command.
/better <idea>) — forge now; one question round only if truly blocking.fast, -f, "just forge it") — never ask. Infer everything, state assumptions in one line above the prompt, forge.run, go) — forge, skip confirmation, execute the forged prompt immediately.If the idea is already detailed, treat it as fast mode even without the flag — there's nothing worth asking.
AskUserQuestion round, real options, only for gaps you can't responsibly default.Include as many of these as the task needs, roughly in this order (stable context first, so it reads top-down and caches well). Each line should leave a reader with no follow-up question.
@path rather than pasting them."Make it good" is not a criterion. Convert it:
@design.png, npm test + typecheck green"If you can't name a checkable bar, you don't yet understand the task — that is the gap worth one question.
Input: /better fast a script to dedupe my newsletter CSV
Assumptions: Python, one CSV, dedupe on email (case-insensitive), keep first occurrence.
Write a Python script that de-duplicates a newsletter CSV.
Input: a CSV with at least an `email` column; take the path as arg 1.
Do: drop rows whose email (lowercased, trimmed) was already seen, keeping the first
occurrence; write the result to <input>-deduped.csv and print "<n> dupes removed".
Constraints: stdlib only (the csv module), don't reorder surviving rows, don't mutate
the input file.
Success: on a file with known dupes it produces a CSV containing exactly the unique
emails, the printed count equals (original rows − unique rows), and re-running it on
the output removes 0. Verify by running it on a small file with a planted duplicate
before saying you're done.
Run it · tweak it · or copy as-is.
For type-specific skeletons (build, fix, refactor, write, research, image, agent, analysis), the reasoning behind each craft rule, and a fuller before→after library, read references/forge-anatomy.md.
Fetches up-to-date documentation from Context7 for libraries and frameworks like React, Next.js, Prisma. Use for setup questions, API references, and code examples.
Applies a firm's KYC/AML rules grid to parsed onboarding records: assigns risk rating, checks required documents, outputs rule outcomes with citations, and routes for escalation.
Generates daily or weekly digests of activity from connected sources (chat, email, docs, tasks, CRM), highlighting action items, decisions, mentions, and project updates.
npx claudepluginhub realdeal88/better --plugin better