ralphctl-minimal-scaffolding

Minimal Scaffolding

"Find the simplest solution possible, and only increase complexity when needed. Every component encodes assumptions about model limitations. Stress-test assumptions; they can go stale quickly as models improve. Remove one component at a time when simplifying. Re-examine the whole system when a new model releases; strip non-load-bearing pieces."

— Anthropic, Harness Design for Long-Running Application Development

Scaffolding built around AI usage — prompts, wrapper scripts, extra validation steps, guardrails — drifts upward in complexity by default. Each piece that solves a real problem at the time of its addition becomes a permanent fixture, even after the model capability that made it necessary has improved past the threshold. Without an active counter-pressure, the scaffolding grows into a weight that slows iteration and obscures the actual design signal. Minimal scaffolding is not a one-time decision at design time; it is a discipline applied on every model upgrade and on every proposed addition.

When this applies

• Refine — before proposing that the requirements need an extra review gate, an extra validation pass, or an approval step to compensate for the model, ask whether the model would produce the right output without it given a well-scoped prompt.
• Plan — before adding an extra AI-assisted review step or splitting the work into more stages, ask whether the additional structure would improve output quality measurably, or whether it is defensive scaffolding against a past model's limitations.
• Execute — before wiring a new wrapper, guard clause, or retry layer around an existing AI-assisted check, ask which assumption about current model capability the addition encodes, and whether that assumption is still valid.

What to do

1. Start with the simplest viable shape. Draft the simplest version that could work given today's model capability. Only add components when the simple version demonstrably fails.
2. Question every component on every model upgrade. When a new model version ships, re-read the project's own notes on why each piece of AI-workflow scaffolding exists, if the project keeps them. For each one, ask: "Would removing this component degrade output quality on the new model?" If the answer is uncertain, run the test.
3. Remove one component at a time, measure. Never remove two pieces of scaffolding simultaneously — you cannot isolate the regression. Remove one; run the project's check gate; observe output quality; decide.
4. Default toward subtraction over addition. When in doubt, omit. Adding a component later when its need is proven is cheaper than carrying a component whose need was assumed.