Distill

Default summarization re-covers the material at lower resolution. This exercises judgment about what matters instead. Two passes, in order — the order is the method.

When to use this

• A doc, diff, research output, or session needs to become something presentable.
• The material is long enough that "what matters" is a real question.
• Skip when the answer is one fact, or when the ask is a faithful in-file summary of stated facts — that is /tldr, which extracts verbatim and never interprets.

Procedure

1. Scope it. The argument, else the material in play this session. Name in one line what you read. Read it fully — a distillation of a skim is worthless.

2. Observation pass. List only the non-obvious:

   • what contradicts the headline or the stated conclusion,
   • what the numbers imply but the text does not say,
   • what is conspicuously absent given what is present,
   • what a skim reads straight past.

   Discard every line that restates the surface. If the list is all surface, say so — "nothing here a skim would miss" is a valid finding, not a failure.

3. Lead pass. Collapse the observations into the 3–4 points you would open with, ranked by what changes the reader's decision — not by what is most interesting or most defensible. Four is the ceiling; three is usually right.

4. Render. Each lead point: one claim-first line, its supporting observation beneath it. Then one line naming what you deliberately left out.

Constraints

• Never run the lead pass first. Selecting before inferring returns the obvious headline — the exact failure this exists to prevent.
• The output is the two lists, nothing else. No preamble, no restatement of the source, no closing recap.
• Length is downstream of selection, not a target. Cutting words while keeping every point is the failure in compressed form.
• An observation you cannot source to the material is a guess — drop it.
• Writes nothing. Chat output only; no artifact, no file edited.

Done criteria

• Observation list contains only things not visible from a skim.
• 3–4 lead points, claim-first, ranked by decision impact.
• What was left out is stated in one line.