Distill

Distill

A summary shortens while keeping proportion — every section of the source gets a smaller version of itself. A distillation does the opposite: it discards most of the source and keeps only what the reader must act on. That trade is deliberate, and it is why a distillation reads shorter and more opinionated than a summary of the same material would.

Reach for this when

• The request is "what do I do about this", "pull the working position out of this", or a corpus is standing between the reader and a decision.
• The source is large enough that quoting or restating all of it would bury the part that matters.
• The output will be read again later, as a live input to work — not filed once and forgotten.

Do not reach for it when the request is "what does this say" or "shorten this" — that is a summary, and proportion is the point of a summary. Do not reach for it to compress a single short document; distillation earns its cost against a corpus, not a paragraph.

The discipline

Five practices, each one a place a distillation quietly fails if skipped.

1. State what was read and what was not. Name every source file consulted, and name what was deliberately skipped and why — a distillation that hides its coverage cannot be audited, and a reader cannot tell "considered and excluded" from "never seen." If a skipped file's content was inherited through another source (a recommendations doc that already quotes it), say that explicitly rather than implying independent verification.
2. Distill, do not quote. State the position in your own words, sourced to where it came from. If checking the output against the source would mostly find matching sentences, it is a merged archive, not a distillation — go back and cut harder.
3. Preserve the provenance of confidence. A number or claim from the source is either what the source measured, or what it inferred, argued, or predicted — carry that distinction into the output instead of flattening everything to flat assertion. An inferred figure that reads as measured is the single most common failure here, and it is always self-flattering: it makes the corpus look more settled than it is.
4. Record disagreement instead of smoothing it. When the corpus and the consuming context's already-decided position diverge, say so and name which one governs. Dropping the conflict to make the output read cleanly destroys the most valuable thing a corpus can hand over — that two sources of judgment did not agree.
5. Say what the output is for. A distillation that will be read again and edited as work continues is shaped differently from one that archives a decision once — name which this is, because it changes what belongs in it.

What this skill is not

It is not a compression technique — cutting word count without changing what is kept is still a summary, just a terser one. It is not an excuse to drop sourcing: every claim in the output still traces to a file, not to the distillation's own authority. And it is not for a corpus you have not actually read — practice 1 only works if there is real coverage to report.

Worked example

docs/loop-engineering.md in Jon Hill's agent-dotfiles repository (merged 2026-08-11) distilled a 23-file external research corpus into one working document for that repository. It names the seven files it read and the sixteen it skipped and why, states positions in its own words rather than quoting the corpus, marks each figure as the source's measurement or its inference, records one place the corpus and the repository's decided position diverged and says which one governs, and states up front that it is a living document future work edits, not an archive. Read it as an example of the shape this discipline produces, not as a template to copy — the divergence it records and the files it names are specific to that corpus and will not recur in a different one.