Digest Skill — Development skill for Claude Code

Digest — compressive editing, not summarization

A Claude skill that turns a long article, paper, or transcript into a **Chinese digest of about one third the length**, condensed sentence by sentence, in the source's own order and voice, opened by a single arguable lens sentence.

It is built against the default behavior of language models when asked to "summarize": pick what looks important, drop the rest, and rewrite in a neutral analytical register. This skill forbids all three.

The rules that make it different

**No deletion by importance.** Every sentence in the source should have a counterpart in the digest. Compression comes from making each sentence shorter, not from choosing which sentences survive. Deciding a block can vanish is the compiler overstepping the reader.

**The denominator is the Chinese character count of a faithful translation**, not the source's word count. For English sources: words × 1.7. Compute it before starting and count again when done. A 19,525-word essay has a denominator of about 33,000 characters and a target of about 11,000. A transcript compresses further, because spoken redundancy is higher.

**Four editing moves, applied to every sentence**: delete empty verbs and measure words; delete degree adverbs and intensifiers; reduce rhetorical inflation to what it actually refers to; turn coordinated clauses into direct subject-predicate.

**Some things are quoted verbatim and never condensed**: pivot sentences that make the argument lock together, the author's own definitions and conclusions, key details inside load-bearing cases, and numbers that anchor magnitude.

**Condensing removes rhetoric; it does not swap it.** No new metaphors, no invented frameworks, no added lyricism. The source's voice is preserved: prose stays prose, analysis stays analysis.

**One lens sentence, and nothing else from the compiler.** The digest opens with a judgment the reader can argue with. All other observation and commentary is deferred to the