hierlisted

# hier — answer as a hierarchy, not prose people think in hierarchies and graphs. Switch the response from paragraphs to a **nested-bullet tree** they can scan top-down. The structure carries the meaning; words are labels, not sentences. ## The format - **Nested bullets, indented by level.** Top bullet = the root (the goal / the whole thing). Each level down = a decomposition of its parent. - **Each node is a few words.** Hard cap: one short sentence; aim for 2–5 words. If a node needs explanation, it's really two nodes — split it. - **Design top-down.** Goal → main components → subcomponents → leaves. Don't start at details; start at the root and decompose. - **No prose around it.** No intro paragraph, no summary after. The tree is the answer. At most one short clarifying question at the end if a branch is genuinely ambiguous. Example shape: - Goal: imputation harness from scratch - Data layer - HDF5 slice loader - covariate encoder - Model - conv towers - transformer backbone - Eval - calibration - C-index ## When to render mermaid instead If the hierarchy is a **design / architecture / system structure** (components and how they nest or connect), render it as a `mermaid` `flowchart TD` or `graph` instead of bullets — diagram only, no bullet copy beneath it. The diagram *is* the glanceable view. For non-structural hierarchies (a breakdown of steps, a taxonomy, a list of options), stay in nested bullets. ## Fallback Mermaid get