Prompt Builder
amart-builder/prompt-builder
Converte brain-dump em prompts de alta qualidade e prontos para rodar
Ver no GitHubSkill para transformar pensamentos desorganizados em prompts otimizados. Clarifica automaticamente, engenheira o prompt com base em pesquisa atual, mostra para aprovação e executa com verificação integrada.
Instalação
npx convez-ai amart-builder/prompt-builder1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Refinar solicitações vagas em instruções precisas para Claude
- Evitar cargo-culting de prompt engineering e focar no que realmente funciona
- Prototipar workflows que requerem múltiplas iterações de prompts
- Capturar intenção do usuário com clarificação seletiva (apenas quando necessário)
Veredito

Skill prático que elimina fricção no ciclo de prompt engineering. Fundamentado em pesquisa (Wharton, Anthropic) e deliberadamente anti-hype, evitando padrões que não agregam valor em modelos modernos como Claude. Excelente para quem precisa transformar ideias brutas em requisições confiáveis.
README
A tradução para o português deste README ainda está na fila.
/prompt: turn a brain-dump into a great result
A Claude skill that takes your messy, unorganized thoughts and turns them into an excellent outcome: clarify → engineer the prompt → show it for your OK → run it → verify.
You don't write the perfect prompt. You brain-dump everything in your head (messy is fine), and the skill builds the clean, complete prompt that modern Claude does its best work from, shows it to you for approval, then runs it for you.
What it does
When you type /prompt and dump your thoughts, the skill:
- Reads the dump and works out the real goal, the deliverable, and the audience.
- Clarifies only if it matters, asking a question only when a missing detail would change the result. A clear dump passes straight through with zero questions.
- Engineers the prompt using current prompt-engineering research, right-sized to the task, with explicit output format, the why behind the request, and an escape clause.
- Shows you the prompt and waits for your OK. It appears in a code block, so you can approve it, tweak it, or save it for reuse before anything runs.
- Runs it and self-checks the result against your original intent before showing you.
Philosophy
This is the anti-cargo-cult version. Based on Anthropic's prompting guidance plus 2025-26 research (incl. Wharton's controlled studies), it deliberately skips the things that don't actually help modern reasoning models:
- ❌ No fake "you are a world-class expert" personas (no reliable accuracy gain)
- ❌ No stakes / tips / threats / "this is CRITICAL" (no benchmark effect)
- ❌ No reflexive "think step by step" (adds latency on models that already reason)
- ✅ Always keeps explicit output-format instructions, context/motivation, and right-sizing
It defaults to the smallest high-signal prompt the task needs, not the longest.
Tuned for Claude Fable 5 (and recent Opus models): a complete spec up front, literal instruction following, trigger conditions for any tools involved, and checkable "done" criteria on build specs.
Install
Claude Cowork
Cowork installs skills from a .skill file you save, not from a local folder. Paste this
into a fresh Cowork chat and Claude will build you that file:
Package a reusable Cowork skill called "prompt" for me, then give me the
installable file so I can save it.
Fetch these two files from GitHub and keep their contents exactly as-is:
- SKILL.md:
https://raw.githubusercontent.com/amart-builder/prompt-builder/main/prompt/SKILL.md
- reference/prompt-engineering.md:
https://raw.githubusercontent.com/amart-builder/prompt-builder/main/prompt/reference/prompt-engineering.md
Build a skill folder with SKILL.md at the top level and the reference file at
reference/prompt-engineering.md (do not edit the file contents. The "name:
prompt" line in SKILL.md is what makes /prompt work). Zip that folder into a
file ending in .skill and present it to me so I get a "Save skill" button.
Do not claim the skill is installed. You can't install it yourself. Just hand
me the .skill file and tell me to click "Save skill," then start a new chat so
/prompt shows up in autocomplete.
Click Save skill on the file Claude gives you, then start a new chat so /prompt
appears in autocomplete.
Claude Code (CLI)
Claude Code loads skills from your local ~/.claude/skills/ folder, so you can install
directly. Paste this into a fresh chat:
Install a custom skill called "prompt" from GitHub:
1. Clone https://github.com/amart-builder/prompt-builder into a temp folder.
2. Create the directory ~/.claude/skills/ if it doesn't exist.
3. Copy the repo's prompt/ folder to ~/.claude/skills/prompt
(so the files end up at ~/.claude/skills/prompt/SKILL.md and
~/.claude/skills/prompt/reference/prompt-engineering.md).
4. Confirm both files are in place and tell me the skill is installed.
After it finishes, start a new chat so /prompt appears in autocomplete.
OpenClaw
The same SKILL.md format works as an OpenClaw skill. Copy the prompt/ directory into
your OpenClaw skills folder.
How to use it
You don't need it every time. Use it when the output really matters.
/prompt then brain-dump everything in your head. Messy is fine.
Example dump:
"ugh i need to email the practice owner who went quiet after our call, keep it warm, remind her we have two other buyers interested but don't sound pushy"
The skill turns that into a clean, complete prompt and writes the email, checking it against your intent before showing you.
What's in here
prompt/
SKILL.md the skill instructions (workflow + rules)
reference/
prompt-engineering.md the technique reference it builds prompts from
Maintained by @amart-builder. Built for the Martin Healthcare Advisors team and anyone who wants higher-quality output from Claude.