The "cut it by a third" editing prompt

Edit this so it is at least a third shorter without losing meaning or
changing my voice.

Rules:
- Cut redundancy, throat-clearing, and qualifiers that do not change the
  claim.
- Do not add anything. No new transitions, no new phrasings that were not
  mine.
- When a sentence survives, keep my wording. This is a cut, not a rewrite.
- If a sentence is doing no work, delete the whole thing instead of
  shrinking it.

After the edit, list what you cut in three buckets: redundancy, hedging,
filler. If any cut changed the meaning, flag it separately.

Text:
{paste your draft}

Ask a model to “make this better” and it returns your writing wearing a suit: smoother, longer, and no longer yours. Every sharp phrase gets rounded, every plain word upgraded to a fancier one. The one thing most drafts actually need, someone ruthless with a red pen, is the one thing the default behavior refuses to be.

So the prompt bans improvement and only permits removal.

Why it works

• “Do not add anything” is the entire trick. Voice does not survive paraphrase, it survives deletion. Restrict the model to cutting and your sentences come back still sounding like you, just fewer of them.
• “Delete the whole thing instead of shrinking it” targets the model’s compulsion to preserve every idea in compressed form. Some sentences are not too long, they are unnecessary, and those are the highest-value cuts.
• The three-bucket cut list is the feedback loop. After a few drafts you see your own pattern in the buckets. Mine is hedging: some version of “at least in my testing” stapled to claims I had already tested.

What good output looks like

Your draft, a third shorter, and when you read it aloud you cannot hear the edit. The flagged-cuts section should almost always be empty; if the model keeps flagging meaning changes, your draft was leaning on the filler more than you thought, which is also worth knowing.

Where it works (and doesn’t)

Claude follows the “no rewriting” constraint most faithfully. GPT-class models do well but occasionally cannot resist swapping a word; the cut list makes those swaps easy to catch. Local 14B models handle pieces under a thousand words fine, but drift back into paraphrase mode on longer ones.

I run it on everything I publish here. The drafts do not get smarter, but they stop wasting your time, which on this site is the actual job.