My automated doubt development process

5 hours ago

The process was developed to regain trust in AI-assisted development after initial over-reliance on LLMs led to issues, emphasizing automated doubt and multi-perspective scrutiny.
Subagents are central to the workflow, specializing in auditing areas that standard AI might miss, using a 'parallax coverage' approach where different vantage points catch different defects.
Phase 1 (Design) starts with AI-generated specifications, followed by a Pre-implementation workflow with agents like Pre-Implementation Architect and Assumption Excavator to validate design, scope, and assumptions, iterating based on findings.
Phase 2 (Development) involves AI building the project with a checklist, avoiding subagents for writes due to trust concerns, and using Post-implementation agents like Code Validator and Security Analyst to audit code quality and security iteratively.
Phase 3 (Wrap-up and Ship) finalizes with a Ship workflow including agents like Anxiety Reader and Release Readiness Validator to ensure readiness for release, requiring multiple iterations for complex projects.
The process acknowledges that quality is subjective with objective goals, relying on operator judgment to terminate iterations, and notes the token cost, making it adaptable from minimal to extensive agent use depending on project needs.
Key tools and examples are shared, with the Assumption Excavator highlighted as universally applicable for auditing assumptions in any artifact.

Hasty Briefsbeta