team-shinchan:micro-execute

To enforce micro-tasks and code reviews during AI-assisted development, break work into 2-3 minute steps with exact file paths and sequential implement-review cycles. This approach uses fresh subagents per task to prevent drift and low-quality commits.

Executing a PROGRESS.md plan into incremental code changes requires generating explicit micro-task plans, dispatching implementer subagents, and running verification commands for each task. Spec compliance and code quality reviewers with retry limits ensure reliable, reviewable outputs.

Two-stage subagent review for micro-task automation works by dispatching a spec compliance reviewer followed by a code quality reviewer for each task. Both stages operate with defined retry limits to prevent hallucination and ensure quality before moving to the next task.

Yes, you can use sequential implement-review loops for CI-driven engineering workflows. This process breaks phases into 2-3 minute concrete tasks, requires capturing verification command outputs, and concludes with a final integration review across all micro-tasks.

AI-assisted coding drifts from project specifications due to unmanaged context expansion. Preventing drift requires time-boxed implementation steps, mandatory spec compliance reviews, and fresh subagents per micro-task to maintain strict alignment with the original plan.

Limitations of micro-task execution include the need to define exact file paths, verification commands, and explicit plans upfront. It requires overhead for small phases and is constrained by defined retry limits for spec and quality reviewers during the implement-review cycle.