AI Engineer Workflow

Problem Addressed

ae-workflow addresses a common failure mode in AI-assisted development: the process drifts when the model starts coding before requirements are clarified, plans cover only part of the scope, tests are not written first, reviews and completion claims lack evidence, and blocked work is retried instead of redesigned. It is aimed at development tasks that need staged, checkable gates.

How It Works

The skill organizes development into a six-gate system: design approval, requirements completeness, test-first execution, two-stage review, evidence-based verification, and stop-on-blocker. Each Gate has an explicit pass condition. For example, Gate 2 requires a complete Requirements Traceability Matrix with no Missing entries, while Gate 5 requires command output evidence rather than claims such as “should” or “probably.”

It also includes a five-loop system: clarification, plan revision, red-green-refactor, review-fix, and completion decision. The workflow is composed through Phase Skills such as brainstorming, writing-plans, executing-plans, subagent-driven-development, and finishing-a-development-branch, with Auto Mode for unattended runs. The material also mentions use with Claude Code or OpenCode + omo, where the latter adds session management, context save/restore, parallel task coordination, and progress tracking.

Boundaries

This skill behaves more like a process constraint layer than a single code-generation feature. Its usefulness depends on whether the required superpowers:* and gsd-* skills are installed and maintained; without those capabilities, the gates and loops may not close cleanly. For one-off small edits, pure documentation tasks, or environments outside Claude Code/OpenCode, running it directly may be heavier than necessary.