📷 Promptograph

Photograph every system prompt that matters & Equip AI Agents with Curated Skills. A production toolkit and FastMCP Server to browse, diff, validate, and generate AI system prompts, built on top of 20,475 real prompts from 55 public repos, plus instant access to 226+ curated engineering skills across 37 domains.

🇧🇷 Português · 🇺🇸 English

---

🇺🇸 English

What is this?

Promptograph is a self-contained toolkit for working with AI system prompts — the hidden instructions that shape how models like Claude, ChatGPT, Gemini, Grok, and others behave.

It was built on top of 20,475 system prompts extracted from 55 public GitHub repositories (plus 6 HuggingFace datasets), totaling over 16.4 million words of real production instructions. The goal is to democratize prompt engineering by giving you the same material the major AI labs use to build their products, and tools to learn from it.

🎯 Why we built this

Most AI users interact with models as black boxes. The system prompt — the hidden "constitution" that defines an AI's personality, capabilities, limitations, and tools — is the most under-discussed piece of the puzzle.

We believe:

• Transparency helps everyone: developers learn faster, researchers find vulnerabilities, users make informed choices.
• Patterns matter more than prompts: the best system prompts share structural patterns. Learning them is more valuable than copying any one prompt.
• Good prompts are engineered, not written: just like code, system prompts benefit from validation, testing, comparison, and iteration.

✨ What makes Promptograph different

There are other "prompt lens" / "prompt scope" projects out there. Here's what they do — and what we do that they don't:

Other projects	Promptograph
Hook into Claude Code for one user	Standalone tool, anyone can use
A/B testing of LLM responses	Diff between real system prompts
LLM-as-judge prompt evaluation	Heuristic validation against 13 best practices extracted from production prompts
Single-tool (just A, or just B)	4 features in one: Browse, Diff, Validate, Generate
Requires API keys, npm, pip	Zero dependencies, runs offline
English only	Bilingual (PT-BR + EN)

🎯 Use cases

For developers: You can build a coding agent in 5 minutes — open Promptograph, filter by "claude coding agent", see how Anthropic structures theirs, customize and ship. No more starting from zero.

For prompt engineers: Before sending a prompt to a client, run it through Validate. Get a 0-100% score with 13 checks and 6 red flags. Like a linter, but for prompts.

For product managers: Diff Cursor vs Windsurf system prompts to understand why one converts better. Diff ChatGPT vs Claude to see how they position their tools internally. Insights you can't get from marketing pages.

For security researchers: Search across 20k+ prompts for "refuse", "harmful", "injection". See how each company handles jailbreak, prompt injection, sensitive content. Real corpus for research.

For researchers and students: A corpus of 20k+ real production system prompts for qualitative analysis. How did the tone change from 2023 to 2026? Which companies added the most guardrails? You can write a paper with this data.

For tech writers and creators: Take a "creative writing" preset, customize it, get a better prompt than 99% of what's out there. Then Validate to check the quality.

For CTOs and tech leads: Compare system prompts to evaluate AI vendors. Which is more conservative? Which has more guardrails? Which is more transparent? Technical decision based on evidence.

✨ Features

📚 Browse

Navigate 20,475+ system prompts with:

• Full-text search (filename, model, persona)
• Filter by company (Anthropic, OpenAI, Google, xAI, Perplexity, etc)
• Sort by size, date, or name
• Per-prompt metadata: model, date, persona, XML tags, tools detected
• Inline preview of raw content

🔀 Diff

Compare any two prompts side-by-side:

• Unified diff with color highlighting (green=added, red=removed)
• Token-level statistics
• Change percentage
• Up to 5,000 diff lines

✅ Validate

Score any system prompt (0-100%, grade A+ to F) against 13 best practices extracted from the most successful production prompts:

• Identity clarity (15 pts)
• Tone guidelines (10 pts)
• Refusals handling (10 pts)
• Safety rules (10 pts)
• Examples section (10 pts)
• Formatting rules (8 pts)
• Tool usage (8 pts)
• Memory/context (6 pts)
• Knowledge cutoff (5 pts)
• Structured tags (5 pts)
• Citation rules (5 pts)
• Limits/boundaries (5 pts)
• Length appropriateness (3 pts)

Plus 6 red flags: vague identity, jailbreak vulnerability, copyright missing, instruction contradictions, too short, too long.

✨ Generate

Build a new system prompt from a spec, using 5 presets based on real production prompts:

• Claude Code-style coding agent (Anthropic)
• ChatGPT 5.5-style assistant (OpenAI)
• Cursor-style IDE agent (Anysphere)
• Perplexity-style search engine (Perplexity AI)
• Devin-style autonomous engineer (Cognition)

Each preset produces a complete, validated prompt that scores 80%+ on the validator.

🏗️ Architecture

promptograph/
├── server.py                  # Python stdlib HTTP server
├── data/
│   ├── index.json             # 5,491 raw prompts indexed
│   └── index_filtered.json    # 20,475 filtered (READMEs removed)
├── generator/
│   ├── parser.py              # Scans repos, extracts metadata
│   ├── builder.py             # Constructs prompts from specs
│   └── refine_index.py        # Filters out non-prompt files
├── validators/
│   └── quality.py             # 13 best practices + 6 red flags
├── static/
│   ├── index.html             # UI
│   └── app.js                 # Frontend (vanilla JS, no build)
├── .github/
│   └── workflows/
│       └── ci.yml             # GitHub Actions CI + Pages
├── bootstrap.sh               # One-command setup
├── publish_to_github.sh       # GitHub publish script
├── LICENSE
├── ATTRIBUTIONS.md
└── README.md (this file)

Stack: Python 3.10+ (stdlib only) + Vanilla HTML/JS. Zero npm, zero pip

🔌 FastMCP Server for AI Agents

Promptograph features a native Model Context Protocol (MCP) server built with FastMCP. Connect it to Claude Desktop, Claude Code, Cursor, Windsurf, or Antigravity to give your AI assistants real-time access to production prompt wisdom and 226+ operational skills blueprints.

MCP Configuration

Add Promptograph to your MCP settings (e.g., claude_desktop_config.json, .cursor/mcp.json or your agent config):

{
  "mcpServers": {
    "promptograph": {
      "command": "python",
      "args": ["scripts/promptograph_mcp_server.py"]
    }
  }
}

MCP Toolset

Tool	Category	Description
promptograph_search	Prompts	Search 20,475 real prompts by company, model, or text
promptograph_validate	Prompts	Score any prompt (0-100%, Grade A+ to F) via 13 heuristic rules
promptograph_generate	Prompts	Generate battle-tested prompts using presets (Claude, GPT, Cursor, Devin)
promptograph_stats	Prompts	Get token counts, company distributions, and model metrics
promptograph_skills_search	Skills	Discover 226+ curated skills across 37 high-leverage domains
promptograph_skills_get	Skills	Retrieve full executable blueprint (SKILL.md) for any skill
promptograph_skills_categories	Skills	Explore the complete breakdown of skill categories

---

🚀 Quick Start

Local (Python)

git clone https://github.com/4pixeltechBR/promptograph.git
cd promptograph
./bootstrap.sh        # Builds the index (one-time, ~2 min)
python3 server.py 8765

Open http://localhost:8765 in your browser.

GitHub Pages (UI only, no backend)

The static UI is in static/. Once you enable GitHub Pages on your fork, the UI will be live at https://4pixeltechBR.github.io/promptograph/. The API calls will fail unless you also deploy the backend (Docker below).

Docker

docker build -t promptograph .
docker run -p 8765:8765 promptograph

📊 What's inside

Metric	Value
Prompts indexed	20,475
Curated Skills indexed	226+
Skill Categories	37 domains (Agents, Audio, Video, Quant, Clean Code, etc)
FastMCP Server	Included (stdio transport)
Tokens indexed	~94,000,000
Words indexed	70.5M
Lines indexed	4.7M
Companies	40+ (Anthropic, OpenAI, Google, xAI, Meta, Mistral, DeepSeek, Moonshot, Zhipu, Cerebras, NVIDIA)
Source repositories	55 (GitHub) + 6 (HuggingFace)
Date range	2022 — 2026
Index size (JSON)	4.8 MB (prompts) + ~380 KB (skills)
Backend RAM	~50 MB
Disk after install	~1.3GB (with archive) or ~25 MB (standalone)

🎯 Use cases

1. Learn prompt engineering — read 5,000+ real examples, identify patterns, build intuition
2. Equip autonomous AI agents — stream curated operational skills and prompt blueprints in real time via FastMCP
3. Build a custom AI agent — start from a preset, tweak, validate
4. Audit your product's prompt — paste it, get a score, see what's missing
5. Compare models — diff Claude Fable 5 vs Claude Opus 4.8 to understand what changed
6. Security research — identify prompt patterns vulnerable to injection
7. Benchmark / regression test — track how a prompt evolves over time

🛠️ API Reference

GET  /api/index                      → All 20,475 indexed prompts (JSON)
GET  /api/raw?id=<id>                → Full content of one prompt
GET  /api/stats                      → Counts and totals
GET  /api/presets/<name>             → Preset spec (claude_coding_agent, etc)

POST /api/diff                       → {left, right} → unified diff
POST /api/validate                   → {content} → score + checks
POST /api/generate                   → {spec} → generated prompt

🤝 Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

📜 License & Ethics

MIT License — see LICENSE.

Source prompts attribution — see ATTRIBUTIONS.md. We respect the licenses of the source repositories and use them for research, education, and transparency.

Ethical use — this tool is for learning, auditing, and building. Do not use it to bypass safety measures, attack production systems, or violate the Terms of Service of any AI provider.

---

🇧🇷 Português

O que é isso?

Promptograph é um toolkit auto-contido para trabalhar com system prompts de IA — as instruções ocultas que moldam o comportamento de modelos como Claude, ChatGPT, Gemini, Grok e outros.

Foi construído em cima de 20.475 system prompts extraídos de 55 repositórios públicos do GitHub, totalizando mais de 12 milhões de tokens de instruções reais de produção. O objetivo é democratizar a engenharia de prompts dando a você o mesmo material que os grandes laboratórios de IA usam para construir seus produtos, e ferramentas para aprender com ele.

🎯 Por que construímos isso

A maioria dos usuários de IA interage com modelos como caixas-pretas. O system prompt — a "constituição" oculta que define a personalidade, capacidades, limitações e tools de uma IA — é a peça mais subestimada do quebra-cabeça.

Acreditamos que:

• Transparência ajuda todo mundo: devs aprendem mais rápido, pesquisadores acham vulnerabilidades, usuários fazem escolhas informadas
• Padrões importam mais que prompts: os melhores system prompts compartilham padrões estruturais. Aprender esses padrões vale mais que copiar qualquer prompt individual
• Bons prompts são engineered, não escritos: assim como código, system prompts se beneficiam de validação, teste, comparação e iteração

✨ O que diferencia o Promptograph

Existem outros projetos "prompt lens" / "prompt scope" por aí. Veja o que eles fazem — e o que a gente faz que eles não fazem:

Outros projetos	Promptograph
Hook no Claude Code pra um usuário	Ferramenta standalone, qualquer um usa
A/B testing de respostas LLM	Diff entre system prompts reais
LLM-como-juiz pra avaliar prompts	Validação heurística contra 13 práticas extraídas de prompts de produção
Single-tool (só A, ou só B)	4 features em uma: Browse, Diff, Validate, Generate
Requer API keys, npm, pip	Zero dependências, roda offline
Só inglês	Bilíngue (PT-BR + EN)

✨ Funcionalidades

📚 Browse

Navegue por 20.475+ system prompts com:

• Busca full-text (nome do arquivo, modelo, persona)
• Filtro por empresa (Anthropic, OpenAI, Google, xAI, Perplexity, etc)
• Ordenação por tamanho, data ou nome
• Metadata por prompt: modelo, data, persona, tags XML, tools detectadas
• Preview inline do conteúdo raw

🔀 Diff

Compare dois prompts lado-a-lado:

• Diff unified com cores (verde=adicionado, vermelho=removido)
• Estatísticas em tokens
• Porcentagem de mudança
• Até 5.000 linhas de diff

✅ Validate

Pontue qualquer system prompt (0-100%, nota A+ a F) contra 13 boas práticas extraídas dos prompts de produção mais bem-sucedidos:

• Clareza de identidade (15 pts)
• Diretrizes de tom (10 pts)
• Tratamento de recusas (10 pts)
• Regras de segurança (10 pts)
• Seção de exemplos (10 pts)
• Regras de formatação (8 pts)
• Uso de tools (8 pts)
• Memória/contexto (6 pts)
• Knowledge cutoff (5 pts)
• Tags estruturadas (5 pts)
• Regras de citação (5 pts)
• Limites/boundaries (5 pts)
• Tamanho apropriado (3 pts)

Mais 6 red flags: identidade vaga, vulnerabilidade a jailbreak, copyright ausente, contradições de instrução, muito curto, muito longo.

✨ Generate

Construa um novo system prompt a partir de uma spec, usando 5 presets baseados em prompts reais de produção:

• Claude Code-style coding agent (Anthropic)
• ChatGPT 5.5-style assistant (OpenAI)
• Cursor-style IDE agent (Anysphere)
• Perplexity-style search engine (Perplexity AI)
• Devin-style autonomous engineer (Cognition)

Cada preset produz um prompt completo e validado que tira 80%+ no validador.

🏗️ Arquitetura

promptograph/
├── server.py                  # Servidor HTTP Python stdlib
├── data/
│   ├── index.json             # 5.491 prompts brutos indexados
│   └── index_filtered.json    # 20.475 filtrados (READMEs removidos)
├── generator/
│   ├── parser.py              # Varre repos, extrai metadata
│   ├── builder.py             # Constrói prompts a partir de specs
│   └── refine_index.py        # Filtra arquivos não-prompt
├── validators/
│   └── quality.py             # 13 boas práticas + 6 red flags
├── static/
│   ├── index.html             # UI
│   └── app.js                 # Frontend (vanilla JS, sem build)
├── .github/
│   └── workflows/
│       └── ci.yml             # GitHub Actions CI + Pages
├── bootstrap.sh               # Setup em um comando
├── publish_to_github.sh       # Script de publish
├── LICENSE
├── ATTRIBUTIONS.md
└── README.md (este arquivo)

🔌 Servidor FastMCP para Agentes de IA

O Promptograph inclui um servidor nativo do Model Context Protocol (MCP) construído com FastMCP. Conecte-o ao Claude Desktop, Claude Code, Cursor, Windsurf ou Antigravity para dar aos seus assistentes e agentes acesso em tempo real à sabedoria de 20.475 system prompts de produção e a mais de 226 blueprints operacionais de skills.

Configuração MCP

Adicione o Promptograph ao arquivo de configuração de MCPs do seu agente (ex: claude_desktop_config.json, .cursor/mcp.json ou similar):

{
  "mcpServers": {
    "promptograph": {
      "command": "python",
      "args": ["scripts/promptograph_mcp_server.py"]
    }
  }
}

Conjunto de Ferramentas MCP

Ferramenta	Categoria	Descrição
promptograph_search	Prompts	Busca em 20.475 prompts reais por empresa, modelo ou texto
promptograph_validate	Prompts	Pontua qualquer prompt (0-100%, Nota A+ a F) com 13 regras heurísticas
promptograph_generate	Prompts	Gera prompts testados usando presets (Claude, GPT, Cursor, Devin)
promptograph_stats	Prompts	Métricas de tokens, distribuição de empresas e modelos
promptograph_skills_search	Skills	Descobre 226+ skills curadas distribuídas em 37 domínios de engenharia
promptograph_skills_get	Skills	Recupera o blueprint operacional executável (SKILL.md) de qualquer skill
promptograph_skills_categories	Skills	Relatório completo de categorias disponíveis e contagens

---

🚀 Quick Start

Local (Python)

git clone https://github.com/4pixeltechBR/promptograph.git
cd promptograph
./bootstrap.sh        # Constrói o índice (uma vez, ~2 min)
python3 server.py 8765

Abra http://localhost:8765 no navegador.

GitHub Pages (só UI, sem backend)

A UI estática está em static/. Quando você ativar GitHub Pages no seu fork, a UI fica em https://4pixeltechBR.github.io/promptograph/. As chamadas de API vão falhar a menos que você faça deploy do backend (Docker abaixo).

Docker

docker build -t promptograph .
docker run -p 8765:8765 promptograph

📊 O que tem dentro

Métrica	Valor
Prompts indexados	20.475
Skills Curadas indexadas	226+
Categorias de Skills	37 domínios (Agentes, Áudio, Vídeo, Quant, Clean Code, etc)
Servidor FastMCP	Incluso (transporte stdio nativo)
Tokens indexados	~94.000.000
Palavras indexadas	70,5M
Empresas	40+ (Anthropic, OpenAI, Google, xAI, Meta, Mistral, DeepSeek, Moonshot, Zhipu, Cerebras, NVIDIA)
Repositórios fonte	55 (GitHub) + 6 (HuggingFace)
Período	2022 — 2026
Tamanho do índice (JSON)	4,8 MB (prompts) + ~380 KB (skills)
RAM do backend	~50 MB
Disco após instalação	~1,3 GB (com archive) ou ~25 MB (standalone)

🎯 Casos de uso

1. Aprender engenharia de prompts — leia 5.000+ exemplos reais, identifique padrões, construa intuição
2. Equipar agentes autônomos de IA — injete skills operacionais e blueprints prontas em tempo real via FastMCP sem inchar o contexto
3. Construir um agente de IA customizado — comece de um preset, ajuste, valide
4. Auditar o prompt do seu produto — cole, receba um score, veja o que falta
5. Comparar modelos — faça diff entre Claude Fable 5 e Claude Opus 4.8 pra entender o que mudou
6. Pesquisa de segurança — identifique padrões de prompt vulneráveis a injeção
7. Benchmark / regression test — acompanhe como um prompt evolui com o tempo

🛠️ Referência da API

GET  /api/index                      → Todos os 20.475 prompts indexados (JSON)
GET  /api/raw?id=<id>                → Conteúdo completo de um prompt
GET  /api/stats                      → Contagens e totais
GET  /api/presets/<nome>             → Spec de um preset

POST /api/diff                       → {left, right} → diff unified
POST /api/validate                   → {content} → score + checks
POST /api/generate                   → {spec} → prompt gerado

🤝 Contribuindo

Contribuições são bem-vindas! Veja CONTRIBUTING.md para diretrizes.

📜 Licença e Ética

MIT License — veja LICENSE.

Atribuição dos prompts fonte — veja ATTRIBUTIONS.md. Respeitamos as licenças dos repositórios fonte e os usamos para pesquisa, educação e transparência.

Uso ético — esta ferramenta é para aprender, auditar e construir. Não use para burlar medidas de segurança, atacar sistemas de produção ou violar os Termos de Serviço de qualquer provedor de IA.

---

🤝 Créditos

Construído por Mavis (MiniMax Agent) com base no trabalho de dezenas de mantenedores de repositórios open source. Veja ATTRIBUTIONS.md.

Maintainer: @4pixeltechBR

📜 License

MIT © 2026