notating-math-naturally

Writes math explanations goal-first, one new symbol per step, with labelled groups. Use to explain a loss, derivation or formula.

SKILL.md

GitHub

Notating math naturally

references/examples.md has a bad and a good passage from a real session, with what makes each work or fail. Read it before writing the first explanation in a session.

Order

  1. Start from what is being optimized or asked, in words. Example: "copy the human driver", not "let L denote…".
  2. Introduce ideas in the order they were devised in real life, each one fixing a problem in the one before. Example: imitation → the flow-model loss that replaces likelihood → fine-tuning with rollouts → AWR → the specific advantage.
  3. No glossary up front. A symbol appears the first time the story needs it, never earlier.

One symbol at a time

  • One display block introduces one new symbol (or one tightly coupled pair, such as a rollout and its index). Say what it is in a plain sentence before the block.
  • After its definition, use only the symbol. Code identifiers (loss_flow, weighted_valid_mask) appear once, where the symbol is defined, and only the symbol after that.
  • Reuse earlier symbols freely; never redefine one. If a quantity changes meaning, give it a new name.

Long expressions

  • Split any expression with more than about three parts into labelled groups using \underbrace{…}_{\text{(1) name}}.
  • Right after the block, say the reading order: "Read (1) first: …; then (2): …". Say what each group contributes, not what it literally computes.
  • When an expression is later specialized (e.g. to one scene, or to one advantage), keep the same group labels so the reader can track which part changed.

Numbers

  • Before a table, work one row in full so the reader can check the method.
  • Label every number as measured, computed from a formula, or estimated. Estimates also say how they were estimated and their known bias.
  • Cite sources (file:line, run IDs, metric names) in prose, not inside equations.