Ray Myers’ “Abstraction Leap” concept suggests designing explicit DSLs rather than letting LLM prompts become source code (source highlight). Bike outlines could be perfect for this: XHTML structure makes them machine-readable while the outliner UI stays human-friendly.

The approach

Template + Validator = Guidance + Guarantees

• Template shows LLMs the expected shape
• Validator enforces that structure after generation
• Result: predictable, testable foundation vs brittle free-form prompts

Connection to specification as code

This aligns with 2.8.4.2: the Bike outline becomes the primary artifact that compounds over time (see 2025-07-19_09-38). Following John Rush’s “fix inputs, not outputs,” you improve the template when plans generate poor breakdowns, not just the individual output.

In practice

Unlike free-form markdown specs that require manual interpretation, Bike’s XML structure makes it easier for LLMs to understand and process. The outliner’s visual hierarchy could make complex plans manageable while maintaining the machine-readable structure needed for reliable AI collaboration. This could bridge human planning intuition with computational precision.