Reduce prompt tokens — trim the words, keep the meaning

Prompts grow by accretion: polite filler, restated context, copy-pasted blocks. Paste your prompt and trim it with live savings per fix — locally.

What bloats prompts the most?

• Filler politeness — "please note that", "I would like you to", "it is important that". Models do not need them.
• Restated context — the same instruction rephrased three times. Once is enough.
• Copy-pasted blocks — repeated examples or previous answers pasted back in.
• Formatting overhead — markdown headers and bullets for content the model only reads once.

Trim without breaking the prompt

Cut words, never constraints. "Answer in German, max 5 bullets, no code" is cheap and load-bearing — keep it. The filler around it is what costs you.

Trimming your side of the conversation is only half the battle — if the context is logs, JSON or stack traces, use the context optimizer which dedupes and compacts data too.