# AI Project Discovery & Architecture Checklist
**Author:** Sebastián Larrauri — Claude Forward Deployed Engineer & AI Systems Architect  
**Practice:** Slackdevs (https://slackdevs.com)  
**Version:** 2.4 (Production Specification)  

---

## 1. Business Feasibility & Use Case Scoping
- [ ] **Problem Definition:** Clear technical statement of the manual/inefficient process being automated.
- [ ] **Quantifiable Success Metrics:** Latency budget (< 1500ms), acceptable error rate (< 1%), token cost ceiling.
- [ ] **Human-in-the-Loop Boundaries:** Identify points where automated execution halts for human sign-off.
- [ ] **Data Sensitivity Classification:** Determine if PII/regulatory data is processed (GDPR, SOC2, HIPAA).

## 2. Model & Ecosystem Architecture
- [ ] **Model Hierarchy Selection:**
  - Fast/Tactical: Claude 3.5 Haiku (routing, extraction, light classification)
  - Workhorse: Claude 3.7 Sonnet (tool use, coding, complex instruction following)
  - Deep Reasoning: Claude 3.5 Opus (strategic synthesis, contract analysis)
- [ ] **Context Window Budgeting:** Establish strict token utilization thresholds and sliding window pruning.
- [ ] **Model Context Protocol (MCP):** Identify external systems requiring standardized MCP connector servers.

## 3. Tool Definition & Deterministic Safeguards
- [ ] **JSON Schema Validation:** Explicit typing, description boundaries, and strict parameter definitions.
- [ ] **Idempotent Execution:** Guarantee tool calls cannot produce duplicated state side-effects.
- [ ] **Zero-Custody Risk:** Externalize financial and sensitive mutations through regulated partner APIs.

## 4. Prompt Engineering & System Prompts
- [ ] **Spec-Driven Prompts:** System prompts maintained as versioned markdown artifacts in repository.
- [ ] **Few-Shot Exemplars:** High-quality production examples covering common edge cases and failures.
- [ ] **Structured Fallbacks:** Graceful error messages and retry mechanisms on JSON validation failures.

## 5. Evaluation, Observability & Deployment
- [ ] **Automated Eval Suites:** Test datasets running against new prompt versions before deployment.
- [ ] **Telemetry Logging:** Log input token counts, output token counts, tool execution times, and cache hit rates.
- [ ] **Prompt Caching Strategy:** Pin reusable system prompts and schema definitions to leverage 90% cost savings.