Iterative Self-Critique

AI plans look reasonable but fail on execution. The model generates a migration plan, an architecture, or a workflow sequence. It reads well. Then you implement it and discover missed dependencies, wrong ordering, or impossible steps. LLMs plan at 50% accuracy on complex tasks.

The Plan-Critique-Revise pattern from Google DeepMind (2025). Three validation steps for every plan: identify preconditions, verify each step satisfies its preconditions, calculate the resulting state. Domain-specific templates for n8n workflows, database migrations, and system architecture. The critique phase catches errors before execution.

Plans that actually work when executed. Accuracy goes from 50% to 90% on planning tasks (DeepMind benchmarks). You catch the broken step before you run it, not after.

• 5 files implementing the Plan-Critique-Revise methodology
• 3 validation steps: preconditions, satisfaction, resulting state
• Domain template: n8n workflow planning and validation
• Domain template: database migration sequencing
• Domain template: system architecture decision validation

• Developers using AI for architecture and migration planning
• n8n builders who need validated workflow sequences
• Anyone who has been burned by an AI plan that looked right but failed