Claude Compass
Vinayak89500/Final_Working_Prototype_NL
Recomendador IA para escolher o melhor modo de execução Claude
Ver no GitHubCamada de inteligência decisória que intercepta a intenção do usuário e recomenda o modo ótimo de execução (Prompt Simples, Workflow ou Agent) antes da execução, com análise de custo de tokens e rationale técnico. Reduz ansiedade decisória e otimiza adoção de capacidades avançadas do Claude.
Instalação
npx convez-ai Vinayak89500/Final_Working_Prototype_NL1 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
- Decidir entre Prompt, Workflow ou Agent antes de executar uma tarefa
- Comparar custos de tokens e qualidade de output entre modos de execução
- Aumentar adoção de Workflows reduzindo fricção decisória
- Decompor tarefas complexas em steps estruturados
Veredito

Prototipo técnico bem estruturado que resolve um problema real de experiência do usuário — a decisão paralisa mais que a execução. Demo ao vivo e stack Python/FastAPI robusto, mas ainda em fase de validação de adoção real e refinamento do motor de classificação.
README
A tradução para o português deste README ainda está na fila.
Claude Compass 🧭
An AI-native decision intelligence layer that helps users decide when to use a Workflow vs. a Simple Prompt vs. an Agent — before execution.
🔗 Live Demo: final-working-prototype-nl.onrender.com
The Problem
As Claude evolved into Skills, Workflows, and Agents, users stopped struggling with what Claude can do — and started struggling with how to use it. ~35% of tasks qualify for a Workflow, but only ~5% of users actually adopt one. The result: token waste, trial-and-error fatigue, and abandoned advanced capabilities.
This isn't a skill gap. It's decision anxiety.
What Compass Does
Compass intercepts user intent before execution and recommends the optimal Claude mode with rationale.
- Task Decomposition — breaks a raw task into Input, Process, and Output steps
- Capability Scorecard — evaluates the task across all Claude execution modes (Simple Prompt / Workflow / Agent) and recommends the best fit
- Token Efficiency Proof — shows side-by-side token cost, output quality, and runtime so users understand why one mode wins
Built For
The NextLeap Product Management Fellowship graduation project (2026 cohort), submitted in the Workflows capability track.
Tech Stack
- Backend: FastAPI (Python)
- Frontend: Vanilla HTML / CSS / JS (single-page)
- AI: OpenRouter API (calling Claude models)
- Hosting: Render
Project Structure
Final_Working_Prototype_NL/
├── index.html # Single-page frontend (Compass UI)
├── main.py # FastAPI backend — handles /analyze and /execute endpoints
├── test_api.py # Decision Intelligence Engine — system prompt + classifier logic
├── requirements.txt # Python dependencies
├── .env.example # Template for environment variables
├── .gitignore
└── README.md
Run Locally
Prerequisites
- Python 3.9+
- An OpenRouter API key
Setup
# 1. Clone the repo
git clone https://github.com/Vinayak89500/Final_Working_Prototype_NL.git
cd Final_Working_Prototype_NL
# 2. Create virtual environment
python -m venv venv
source venv/bin/activate # macOS / Linux
# OR
venv\Scripts\activate # Windows
# 3. Install dependencies
pip install -r requirements.txt
# 4. Set up environment variables
cp .env.example .env
# Then open .env and add your OPENROUTER_API_KEY
# 5. Run the server
uvicorn main:app --reload
# 6. Open your browser
# http://localhost:8000
How It Works
User input
↓
Frontend (index.html) — captures task description
↓
POST /analyze → main.py
↓
test_api.py — Decision Intelligence Engine
• Step-count parser (verbs, "then", lists)
• Recurrence detector (weekly, every, recurring)
• Determinism scorer (structured vs creative output)
• Token estimator + confidence gate (<70% → ask 1 clarifying question)
↓
JSON Scorecard returned to frontend
• Recommended mode + rationale
• Token cost across all 3 modes
• Step-by-step decomposition
↓
User reviews → executes → rates result
Project Context
This repository is the prototype layer of a complete product strategy. The full project includes:
- 📑 10-slide pitch deck — problem, research, personas, solution, GTM, metrics
- 🎙️ User research — in-depth interviews + survey of Claude power users
- 📈 Product strategy — North Star metric (Workflow Activation Rate), four-signal classifier, embedded in-product GTM
- 🎬 90-second demo video
What I'd Do Differently
A few things I'd change if taking this from prototype to production:
- Larger sample size for research. The qualitative signal was strong, but the survey size was exploratory. A v2 would target n=200+ with stratified sampling across user behaviour segments.
- Effort estimate was optimistic. I scoped this at 4–6 weeks. A real implementation — with proper classifier ML, persistence, and Anthropic plan-tier integration — is closer to 12–16 weeks for a small team.
- The classifier is rule-based, not learned. The four-signal heuristic works for the demo, but a production version should fine-tune on labelled task → optimal-mode pairs as users interact.
- No team mode yet. The "Reluctant Team Coordinator" persona was identified but not solved in MVP. A shared Workflow library is the natural v2.
Built By
Vinayak — aspiring APM, NextLeap Fellowship 2026.
Built as part of the NextLeap PM Fellowship graduation requirement. Not affiliated with Anthropic.