terse

Strip boilerplate from LLM conversation history before sending it back to the model. Less noise, fewer tokens, same signal.

"Certainly! I'd be happy to help. In order to fix this issue,
 you should take into consideration the edge cases. I hope this helps!"

→  "Fix this issue, consider the edge cases."

  30 tokens → 12 tokens  (-60%)

Install

curl (macOS and Linux):

curl -fsSL https://raw.githubusercontent.com/fwdai/terse/main/install.sh | bash

cargo:

cargo install terse-rs

Manual: download a pre-built binary from Releases, extract, and place terse on your $PATH.

Packages

Package	Language	Description
packages/terse	TypeScript (Bun)	npm library + CLI
crates/terse	Rust	library + binary

Shared test fixtures live in fixtures/ — both implementations are verified against the same cases.

Testing

make test        # run both suites
make test-ts     # TypeScript only (bun test)
make test-rust   # Rust only (cargo test)

What it does

LLM conversations accumulate filler: opener affirmations (Certainly! Great question!), closer offers (Let me know if you have any questions), verbose phrasing (due to the fact that → because), and structural announcements (Here is the solution:). None of this helps the model on the next turn.

Terse removes it — deterministically, in microseconds, with no external calls.

Three tiers, applied in order:

Tier	What it does	Requires
rules	Regex patterns: boilerplate openers/closers, phrase substitutions, filler words, structural labels	nothing
nlp	POS-aware: drop articles, shorten synonyms (utilize → use, repository → repo)	bun add compromise (TS)
llm	Semantic rewrite via local model	planned

Code blocks, inline code, and URLs are never touched.

Typical savings

Content type	Rules only	Rules + NLP
Conversational (polite, verbose)	40–60%	55–75%
Technical (code-heavy, terse)	5–15%	10–20%

Notes

• English only. No language detection — other languages pass through unchanged.
• Target is the model, not humans. Output is optimized for token efficiency, not readability.
• No semantic loss measurement. Works well in practice; no theoretical guarantee.