Skip to main content

​
Overview

Requirements First is the most comprehensive workflow. It starts by gathering requirements and generating a Product Requirements Document (PRD), then moves through a technical specification, implementation plan, and staged implementation with AI code review. Designed for larger efforts where stakeholder alignment matters.
Credit Efficiency: Requirements First has a higher credit cost (full PRD + spec + implementation + review). Learn more in the Workflow Credit Efficiency reference guide.

​
When to Use

  • Large features spanning multiple files or services
  • Cross-team initiatives needing stakeholder alignment
  • Work requiring formal documentation (PRD, spec)
  • Regulated environments where traceability is important
  • When you want AI code review at every stage

​
How It Works

1

Requirements Gathering

The agent creates a PRD (requirements.md) based on your task description. This captures scope, user stories, acceptance criteria, and constraints.
2

Technical Specification

Based on the approved requirements, the agent produces a detailed technical spec (spec.md) covering architecture, dependencies, and verification strategy.
3

Implementation Plan

The agent breaks the spec into ordered implementation steps in plan.md, each representing a coherent, testable milestone.
4

Staged Implementation

The agent implements step by step. AI verifier agents run after each stage — executing tests, lint, and spec validation to catch issues before they compound.

​
Artifacts

​
AI Code Review

Requirements First includes automated verification between implementation stages:
  • Verifier agents run tests, lint, and security checks after each step
  • Failures are caught before code reaches the next phase
  • Diffs are reviewed continuously during implementation

​
Tips

Share requirements.md with product managers or team leads before moving to the technical spec. This is the cheapest place to catch scope issues.
Disable Auto-start to pause between phases. Review requirements.md, then spec.md, then approve implementation. Each checkpoint reduces rework.
Use a reasoning model (Opus, GPT-5.4) for requirements and spec, a fast model (Gemini Flash, Sonnet) for implementation, and a different model for review. See Multi-model.

All Workflows

Compare all available workflows

Spec First

Lighter alternative without the PRD phase

Multi-model

Optimize costs with different models per phase