Prompt Smith
hstefanov1/prompt-smith
Engenharia e avaliação sistemática de prompts com versionamento integrado
Ver no GitHubToolkit completo para engenharia, versionamento e avaliação de prompts com Claude. Permite definir objetivos de sucesso para seus prompts, iterar versões e rastrear resultados de forma estruturada através de arquivos JSON.
Instalação
npx convez-ai hstefanov1/prompt-smith1 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
- Otimizar prompts através de iterações versionadas com métricas claras
- Definir e medir objetivos de sucesso para diferentes casos de uso com Claude
- Rastrear histórico completo de versões de prompts e seus resultados
Veredito

Ferramenta prática para quem trabalha intensivamente com prompts e precisa de rigor na iteração e avaliação. Abordagem estruturada com versionamento automático e rastreabilidade clara entre prompts e resultados, sem overhead desnecessário.
README
A tradução para o português deste README ainda está na fila.
PROMPT-SMITH
A toolkit for engineering, versioning, and evaluating prompts with Claude.
Requirements
- Bun v1.3 or higher — https://bun.sh/
- Anthropic API key — https://console.anthropic.com/
Quick Start
bun run verify # Install dependencies and run tests
prompts/
Each file in this directory is a .json that contains the prompt itself and the list of goals it must achieve. This is your starting point — define your prompt and what success looks like before running any eval.
Each iteration of a prompt is saved as a new file (v1, v2, v3...), with the version also reflected inside the file. The goals stay the same across versions. Only the prompt changes as you refine it.
{
"version": 1,
"prompt": "Explain quantum computing in simple terms.",
"goals": [
{ "goal": "The response must be under 100 words" },
{ "goal": "The response must not use technical jargon" }
]
}
results/
Each file in this directory is a .json that contains the result of an eval run. Each result file mirrors its prompt version by name (e.g. prompts/quantum-explanation-v1.json → results/quantum-explanation-v1.json), so you can always trace back which prompt produced which result and compare them over time.
{
"version": 1,
"prompt": "Explain quantum computing in simple terms.",
"score": 6,
"passed": false,
"goals": [
{ "goal": "The response must be under 100 words", "score": 8 },
{ "goal": "The response must not use technical jargon", "score": 4 }
]
}
Usage
Add your prompt file to prompts/, then run the suite by passing the prompt filename as the [suite] argument.
If [suite] is not provided, the script will prompt you to choose from available suites.
./smith-run.sh # Prompt you to choose a suite to run
./smith-debug.sh # Prompt you to choose a suite to debug
./smith-run.sh [suite] # Run a specific suite
./smith-debug.sh [suite] # Debug a specific suite
Example:
./smith-run.sh quantum-explanation-v1.json
./smith-debug.sh quantum-explanation-v1.json
License
MIT — free to use, modify, and distribute.