The Self-Improving Loop

GDD is designed to evolve through use, not just through upfront design. The framework starts minimal and grows through a continuous feedback cycle.

The Cycle

graph LR
    A["Sessions produce<br/>observations"] --> B["Thalamus<br/>captures them"]
    B --> C["Housekeeping<br/>reviews and promotes"]
    C --> D["Skills and instructions<br/>get updated"]
    D --> E["Agent behavior<br/>changes"]
    E --> A

What Evolves

Each housekeeping audit is an opportunity to refine the framework:

• Recurring friction becomes a new skill or ws subcommand
• Validated preferences become committed instructions
• Stale concerns get pruned
• Capture heuristics get tuned — "we're capturing too much noise about formatting, let's stop noting those"

Co-Evolving Artifacts

Three artifacts evolve together:

Artifact	What it does	How it evolves
Thalamus.md	Captures observations	Sections added/removed based on what's useful
Orientation skill	Governs capture behavior	Heuristics tuned by housekeeping feedback
Housekeeping skill	Audits and promotes	Process refined as we learn what promotions are valuable

None of these is "done." Each audit cycle refines the template, adjusts capture heuristics, or adds new sections to the file.

The Principle

Evolve through use — the framework refines itself through audit cycles; no part of it is final.

This is GDD's 6th design principle. It means the initial implementation is deliberately minimal. Rather than trying to design the perfect framework upfront, GDD provides just enough structure to start capturing observations, then lets real usage reveal what's worth formalizing.

A framework that can't change is a framework that will be abandoned. GDD is designed to be the opposite — it gets better the more you use it.