Why Spec-Driven Discovery Matters
Building AI features without upfront technical discovery inevitably results in architectural failure modes: unpredictable token costs, non-deterministic API failures, and regulatory compliance breaches.
This checklist guides engineering teams and product leaders through the five critical phases of architectural vetting before committing engineering resources.
The 5 Discovery Phases in the Checklist
1. Business Feasibility & Scoping
- Defining the automated workflow with exact input and output bounds.
- Establishing latency budgets (< 1500ms) and token expense ceilings per transaction.
- Defining Human-in-the-Loop escalation triggers and approval boundaries.
2. Model Hierarchy & Provider Architecture
- Selecting the optimal model tier (Haiku for routing, Sonnet for tool execution, Opus for complex synthesis).
- Designing sliding context window strategies to avoid state degradation.
- Formulating prompt caching breakpoints on system instructions and tool definitions.
3. Tool Definitions & Zero-Custody Risk
- Specifying typed JSON Schemas for every tool argument.
- Enforcing idempotency to prevent duplicated financial or state mutations.
- Externalizing transaction execution to certified partner APIs.
4. Prompt Engineering & System Specs
- Maintaining system prompts as version-controlled markdown artifacts.
- Supplying few-shot exemplars covering negative and edge conditions.
- Implementing schema validation fallbacks on unstructured output parsing.
5. Observability & Continuous Evaluation
- Configuring automated regression test suites in CI/CD before prompt release.
- Logging token counts, cache hit rates, tool latencies, and user feedback signals.
Enterprise Consulting Integration
Need assistance executing a formal AI Discovery for your team? Sebastián Larrauri conducts structured 1-to-2 week Discovery Sprints to audit architecture, validate feasibility, and establish production roadmaps.
