prompt-trim

Losslessly shorten LLM prompts. Drop-in middleware + CLI. Zero dependencies, idempotent, transform-by-transform opt-out.

npx prompt-trim request.json --report > trimmed.json
# prompt-trim: 14820 -> 12430 tokens (-2390, -16%) | 4 dup paragraphs | 7 tools trimmed

What it does

| transform | confidence | description | |-----------|------------|-------------| | collapse-blank-lines | high | runs of 3+ blank lines -> 2 | | trim-trailing-whitespace | high | strip trailing spaces/tabs per line | | trim-leading-trailing-doc | high | strip leading/trailing whitespace from message | | collapse-spaces | high | runs of 3+ spaces inside a line -> 1 (skips code fences) | | normalize-line-endings | high | CRLF/CR -> LF | | compact-list-bullets | high | * and + bullets -> - | | strip-html-comments | high | drop <!-- ... --> | | compress-horizontal-rules | high | normalize ***, ___, etc. -> --- | | reduce-md-headings | medium | #### and deeper -> **bold** | | replace-verbose-fillers | medium | "in order to" -> "to", strip "please make sure to" | | dedup-paragraphs | high | drop verbatim paragraphs after first occurrence across messages | | trim-tools | high | drop additionalProperties: false, empty required: [], $schema, schema examples; collapse description whitespace |

Confidence levels:

• high runs at every level. Tested across GPT-4o, Claude 3.5 Sonnet/Haiku, Gemini 1.5/2.5.
• medium runs at the default medium level. Spot-checked but changes the visible markup style.
• Aggressive level is reserved for future opt-in transforms; it currently behaves the same as medium.

Code blocks (``` fenced and `single-backtick`) are preserved verbatim by every transform that touches text content.

Install

npm i prompt-trim
# or as CLI
npx prompt-trim request.json

CLI

prompt-trim request.json --report > trimmed.json
prompt-trim request.json --in-place
cat req.json | prompt-trim --stdin --json-report
prompt-trim request.json --level safe          # high-confidence only
prompt-trim request.json --aggressive          # all transforms
prompt-trim request.json --disable replace-verbose-fillers
prompt-trim request.json --enable-only collapse-blank-lines,strip-html-comments
prompt-trim request.json --no-dedup

Library

const { trim } = require('prompt-trim');

const result = trim({
  messages: [
    { role: 'system', content: '   You are helpful.\n\n\n\n   Respond in JSON.   ' },
    { role: 'user', content: 'hello' },
  ],
  tools: [],
  level: 'medium',
});

// result.messages          - shortened messages
// result.tools             - shortened tool defs
// result.stats.tokensBefore
// result.stats.tokensAfter
// result.stats.percentSaved
// result.stats.duplicateParagraphsDropped
// result.stats.toolsTrimmed

As OpenAI middleware

const OpenAI = require('openai');
const { trim } = require('prompt-trim');

const client = new OpenAI();

async function call(body) {
  const trimmed = trim(body);
  return client.chat.completions.create({ ...body, ...trimmed });
}

As Anthropic middleware

const Anthropic = require('@anthropic-ai/sdk');
const { trim } = require('prompt-trim');

const client = new Anthropic();

async function call(body) {
  const t = trim({ messages: body.messages, system: body.system, tools: body.tools });
  return client.messages.create({ ...body, messages: t.messages, tools: t.tools });
}

Idempotency

const { trim, isIdempotent } = require('prompt-trim');
isIdempotent({ messages: [...] }); // true - safe to apply more than once

This matters for retries, replays, and caches: trimming the same input twice is guaranteed equivalent to trimming it once.

Companion tools

• prompt-medic - audit first to see where the waste is.
• prompt-cache-key - structure prompts so OpenAI / Anthropic prompt caching actually hits.

License

MIT.