Brevity

Brevity: every word loaded with meaning

An editing pass that compresses a supplied text to its load-bearing minimum: remove every word that does not change meaning, and no more.

The hard part is the second clause. The value is protecting words that look decorative but are load-bearing: a hedge, a scope, a number, a turn of voice. Cut filler, keep meaning. A cut that drops meaning is deletion, not brevity.

Invocation

Explicit only. Engage only when the user names the skill directly: "use brevity on this", "run brevity on this file", "brevity this", or the /brevity command. Do not auto-trigger on generic requests to shorten, tighten or trim text, and never apply this to your own conversational replies. It operates only on a text the user hands you.

Scope

• In: any supplied text, pasted or a file path you read. Any language. Prose and short-form alike (commits, PRs, Slack, emails, summaries, docs, READMEs, reports).
• Out: fetching URLs, code and your own chat output. Text only.

Procedure

1. Ingest: take the pasted text, or read the file path
2. Classify register → set the compression ceiling (below). State the call in one line
3. Build the ledger: internally extract the load-bearing atoms the rewrite must preserve. Terse, not narrated. See references/taxonomy.md
4. Cut filler up to the ceiling. See references/taxonomy.md
5. Fidelity check: verify every ledger atom survives. A dropped atom is restored, or surfaced in the cut-list for the user to confirm
6. Deliver: rewrite + at-risk cut-list + word-count delta
7. Offer the dial: invite harder or softer, per-section if useful

Register → compression ceiling

Register sets an automatic ceiling on how hard you may cut. Always on, not a user choice. Infer it from signals, state it in one line, proceed. Ask only when genuinely ambiguous.

"Aggressive" never crosses the fidelity line. "Gentle" means tone is content here, not filler.

Aggressiveness

Register sets the ceiling. Within it, default to an assertive, fidelity-safe cut. The shown cut-list is what licenses boldness: nothing is lost silently, so the user can restore.

Optional intensity shortcuts, dialed in conversation, not required upfront:

• gentle: conservative, light touch
• standard: the assertive default
• ungaretti: maximal, still fidelity-bound

The user steers per-section when useful ("full Ungaretti on the intro, softer elsewhere").

Fidelity check + clarity override + "nothing to do"

Fidelity check. After cutting, scan: did any ledger atom vanish? A targeted scan, not a fresh re-derivation. If gone, restore it or surface it as a deliberate removal. The ceiling on aggression is this check: cut until the next cut would drop an atom. Rigor scales with stakes: short-form gets an internal read-back; prose gets an explicit ledger and a shown cut-list; very long docs are chunked by section so the ledger is never held whole.

Clarity override. Some content stays full even when words could be cut, because compressing it harms comprehension or safety. Back off the ceiling for: security and risk warnings, irreversible-action confirmations and multi-step instructions where terseness invites misreading. Fidelity protects claims; this protects clarity. When a cut would create ambiguity, don't make it.

Nothing to do. If the text is already lean, say so ("already tight; only trivial trims available") and stop. Word count is a consequence of removing filler, never a target. Never invent or pad cuts to manufacture a reduction. A skill that always "achieves" a cut will damage good text.

Verbatim-preserve list (keep the exact characters)

Distinct from the ledger. The ledger keeps meaning; this keeps the literal text. Never paraphrase or "tighten" these, even when wordy:

• Quotes (anything the author is quoting)
• Error messages, exactly as written
• Code, commands, anything in backticks
• Identifiers, function and variable names, file paths
• Numbers, versions, dates
• Proper nouns, citations, references
• Defined terms (a term the text explicitly defines)

Conflating the two rules corrupts a citation or rewrites a quote.

Punctuation

Marks face the same test as words: keep what changes meaning, cut what only decorates. A mark is filler when the sentence reads the same without it.

Examples:

• Em dashes: usually a decorative pause. Use a colon, comma, period or nothing, unless the dash sets off a genuine aside or carries the author's voice
• Serial (Oxford) commas: cut where no ambiguity follows, keep the one case where dropping it merges two list items
• Comma splices and stacked clauses: split into shorter sentences, or drop the clause
• Scare quotes, exclamation marks, parenthetical asides: keep only when they change how the claim reads

Load-bearing punctuation stays: a comma that disambiguates, voice-carrying marks and the punctuation inside verbatim-preserve content, which is never altered.

Emission discipline

Output tokens are expensive; the naive build emits the text three or four times (original, ledger, rewrite, diff). Build the cheap version:

• Never re-print the source. It is already in context.
• Show cuts, not copies. The change summary names what was removed, not two full versions side by side.
• Keep the ledger internal and terse. Reasoning scratch, mostly unemitted. Apply Ungaretti to the ledger too.
• Summarize the bulk, enumerate only the risky. Describe routine filler removal in grouped terms ("cut throat-clearing throughout"), never word-by-word; spell out individually only the at-risk cuts (a dropped qualifier, scope or voice) so they can be confirmed.

Output format

The rewrite is shown only for pasted input. For a file edit the diff is the rewrite, so never reprint it; emit the change summary, delta and dial only. Count the delta exactly from both versions (e.g. wc -w), never estimate.

• Short-form: the rewrite (pasted) or a one-line note (file). Fold a terse "what changed" into the note ("cut throat-clearing and hedge-stacking, −22%"). No visible ledger
• Prose / high-stakes:
  • the rewrite, for pasted input only;
  • a short "what changed" summary, grouped by kind, not a per-word log ("removed throat-clearing and hedge-stacking; collapsed two rule-of-three lists; cut ceremonial transitions");
  • any at-risk cut called out separately with a one-clause confirm-this reason ("dropped 'in most cases': confirm the claim is still meant to be hedged");
  • word-count delta (820 → 540 words, −34%), reported as a result, never as a goal hit;
  • the dial offer.

Delivery

Invocation is explicit and name-gated, so act directly. The host's edit-permission prompt and diff view are the confirmation and preview layer, so do not add a bespoke confirm step.

• Pasted text → return the rewrite in chat. It is the deliverable.
• File path → edit the file in place. The diff is the rewrite, so never reprint the rewritten text; surface only the cut-list and delta. If the user asks to preview without changing the file, return the rewrite in chat instead.

Voice

Preserve the author's voice by default: it is load-bearing unless also redundant. Offer a neutralize mode (plain dense prose, voice removed) only when asked. Never neutralize silently.

Anti-patterns

• Padding or inventing cuts to hit a word count
• Refusing to return "already lean" when the text is genuinely tight
• Paraphrasing anything on the verbatim-preserve list
• Dropping a hedge, qualifier or scope and calling it concision
• Re-printing the full original alongside the rewrite
• Narrating the full ledger to the user
• Invoking or depending on another skill: brevity runs standalone
• Touching your own chat replies or running unprompted
• Sanding off the author's voice without being asked
• Dropping articles, breaking grammar, reducing prose to telegraphese. Concision preserves grammaticality and register. This is the line between brevity and a blunt token-stripper. Output must read as natural, well-edited writing

For the full filler and protect taxonomies see references/taxonomy.md; for worked before/after pairs see references/examples.md.