workflow-execute-plans

The workflow-execute-plans skill provides a robust, iterative framework for AI agents to translate complex implementation plans into production-ready code. By enforcing a 'batch execution' strategy, it ensures that an agent does not attempt too much at once, which reduces the risk of logic drift, environment configuration errors, or unverified changes. It is designed for developers who need to maintain strict oversight over AI-driven development tasks, particularly for multi-step feature implementation or refactoring projects.

The skill operates by reading a defined plan, performing a critical review for potential dependencies or risks, and then executing in batches of three tasks. After each batch, the agent automatically provides verification evidence (such as test results, type checks, or linting outputs) and triggers a pause for human review. This cycle repeats until the plan is complete, culminating in a final summary and optional cleanup tasks.

• Performs automated critical plan reviews to identify missing dependencies, incorrect task ordering, or insufficient acceptance criteria.

• Implements logic in safe, default 3-task batches to ensure high granularity and easy rollback.

• Enforces strict verification after each batch including build, lint, and test execution.

• Implements mandatory approval gates: pauses after every batch and when blocked, preventing autonomous drift.

• Maintains persistent execution state in logs/state.json or markdown files to allow for safe interruption and resumption.

• Supports optional skill-evolution loops, suggesting patches to improve the agent's performance based on the current run's success or failure.

• Best for: Complex multi-file refactoring, new feature implementation, and infrastructure-heavy development tasks.

• Inputs: Requires a path to a plan file (typically in 03-plans/), repository root, and a working run directory.

• Outputs: Generates detailed execution logs, per-batch verification reports, and a final 05-final/summary.md for audit.

• Operational constraints: Agents are instructed to stop immediately if they encounter blocked states, failing tests, or ambiguous instructions, prioritizing reliability over speed.