Multi-Agent Collaborative Development Workflow

Problem

When a fuzzy requirement, bug, diff, or technical debt is handed directly to AI, the common failure mode is scope creep: the same context analyzes and edits at once, refactors by habit, mixes multiple goals into one task, and becomes hard to verify. This skill addresses how to keep human decision-making intact in multi-agent development, while assigning AI only clear, bounded responsibilities in decomposition, execution, and review.

How It Works

The skill separates the workflow into an orchestration phase and an execution phase, even when only one AI is used.

• Classify the task first: new requirements, bugs, execution results, technical debt, data/config changes, and documentation reuse get different orchestration paths.
• Size the work: XS can become an atomic task, M needs a proposal first, and L must be split rather than delivered in one shot.
• Constrain atomic tasks: include background, allowed files/modules, forbidden scope, exact fields/functions, acceptance criteria, and self-test commands.
• Review in order: check scope, unrelated refactoring, main flows, data/config/permissions, compatibility, edge cases, and verification before closing.
• Capture reusable assets: maintain files such as AGENTS.md, TASK_TEMPLATE.md, and REVIEW_CHECKLIST.md for long-term projects.

Boundaries And Notes

It fits development collaboration where scope control and verifiable output matter, not open-ended exploration. Do not ask execution agents to decide the best approach, do not pack multiple requirements into one task, and do not mix technical debt with business changes. The material stresses real acceptance over automated checks, and no large edits before proposal confirmation.