Four hours. That's how long Marcin Ostrowski, CTO of fryga, says it takes to go from concept to production on a medium-to-large Rails feature using his custom harness for Claude Code.
The breakdown: one hour of planning, two hours of Claude doing the actual coding, and one hour of QA, review, and deployment. The developer's job shifts from writing code to designing constraints.
Ostrowski's harness, built on Jesse Vincent's superpowers framework, treats AI agents as "hands" that follow predefined rules rather than making independent architectural decisions. The core problem it solves is coordination. Without structure, multiple Claude agents working on the same codebase will independently make conflicting choices about patterns, naming, and architecture. Speed without guardrails just creates technical debt faster.
The system works in three phases. First, a brainstorming dialogue with Claude produces design documents that capture the chosen approach and trade-offs. Second, an implementation plan breaks work into 2-5 minute tasks following test-driven development, with instructions written at the intent level (what to do and where) rather than as exact code snippets. Third, execution runs through automated reviews checking specification compliance, code quality, and Rails conventions, with local CI tools like RuboCop validating each step.
Hooks enforce convention loading before any file edits, and cleanup agents scan for deviations through automated refactoring pull requests. It's an opinionated setup that trades flexibility for consistency.
This pattern of wrapping AI coding tools in project-specific constraints is showing up more often as teams move past the "let the AI figure it out" phase. The interesting part isn't the speed claim. It's the acknowledgment that autonomous AI coding only works when you spend real time upfront defining what "good" looks like for your specific project.