x-ray — /refine: the abstraction & nudge topology

Mode: artifact · Renderer: L0 HTML/CSS layered cascade (two views, no remote deps) · Target: .claude/skills/refine/SKILL.md (v0.2.0). One skill, sliced into the — strata of language it stacks to steer an LLM.

A skill file is not one instruction — it's a stack of control layers. Each layer uses a different linguistic device (a "nudge") operating at a different altitude of abstraction: a binding harness contract at the floor, an invariant skeleton above it, discretionary branching higher still, hard stop-gates, proof obligations, and finally the output shape. Toggle, isolate, or trace a reasoning moment to see which layers fire and in what order. Then switch to Calls · loop for the other dimension — what refine does: the 10-stage canonical loop where it delegates to invoke (define · design · plan), interrogation, distill, and context-builder.

Layer 1 / 13

you’re building the stack one layer at a time · use ← → keys

Drill-in · inside Layer 6

Conditional overlays — a stack inside the stack

Source-backed: every quote and section name is lifted from SKILL.md. Inference: the "device" labels and the "how it nudges" column are my reading of why the language is shaped that way — they are interpretation, not claims the skill makes about itself. Trace orderings are illustrative paths, not a guaranteed execution order.