Rewrite Prompt Generator

The instruction is the ceiling

A rewrite is only as good as the brief behind it. Hand a model "make this better" and it guesses — usually longer, more formal, padded with connective tissue you never asked for. The model is not failing. It has no target, so it defaults to the average of everything it has seen.

Everything that makes a rewrite good lives in the instruction. Four decisions:

• Tone — what the text should sound like.
• Fidelity — what must not change.
• Shape — how long, how structured.
• Delivery — what comes back, and nothing else.

Get those explicit and the rewrite usually lands on the first try. The prompt-writing guide applies the same structure to generation rather than revision.

Name the tone in words the model can act on

"Professional" is not a tone, it is a mood board. It covers a legal brief, a LinkedIn post, and a hospital discharge letter, which share nothing. Vague tone labels are the biggest cause of disappointing rewrites.

Name behaviors instead of adjectives:

• Instead of *friendly*: "use contractions, address the reader as you, one idea per sentence."
• Instead of *professional*: "no exclamation marks, no hedging, lead each paragraph with its conclusion."
• Instead of *concise*: "cut every sentence that adds no new fact; target 40% shorter."
• Instead of *confident*: "remove 'I think', 'perhaps', 'it seems'; state claims directly."

Better still, show it. Paste two or three sentences that already sound right and say "match this voice." A sample beats any adjective list, which is why few-shot prompting wins here. If you write in a house voice, park that sample in a reusable system prompt so every rewrite starts from it.

Protect the facts, control the length

Rewrites drift. Numbers get rounded, "up to 30%" becomes "30%", and a legally load-bearing caveat vanishes because it read awkwardly. Left unsaid, the model treats every word as negotiable.

State the fidelity rules directly:

• Keep all numbers, dates, names, and units exactly as written.
• Add no claims, examples, or statistics that are not in the source.
• Do not remove qualifiers like "approximately", "in most cases", "may".
• If something is ambiguous, keep the original wording rather than guessing.

Length needs the same treatment. "Shorter" gets you 5% shorter. Give a number and a unit: "between 90 and 120 words", "at most three sentences". Ranges beat exact counts, because the model pads to hit a figure. Check output against the token counter when length is tied to a hard limit.

Models also love to introduce themselves: "Sure! Here's a revised version:" then the text, then "Let me know if you'd like changes!" If you are pasting into a CMS, that wrapper is friction. Close it off:

Save briefs that work in the prompt library, and run new ones through the analyzer if you suspect an instruction is ambiguous.

Rewriting for a different reader

Tone and audience are separate dials. You can write formally for a specialist or formally for a nervous first-time customer, and the vocabulary changes completely.

When the audience shifts, say who they are, what they know, and what they do next:

Name the reading level as a number or grade band — "simple" is as slippery as "professional". Say what to do with jargon too: define, replace, or drop. Those give three different documents.

Before and after

Source paragraph, the kind that shows up in internal docs everywhere:

It should be noted that the utilization of the new reporting dashboard has been made available to all users as of the current date, and it is anticipated that the aforementioned tool will facilitate an improvement in the efficiency of the monthly reporting process by approximately 25%, subject to individual team workflows.

Brief: direct tone, second person, keep the 25% and the qualifier, under 45 words, return only the text.

Result:

The new reporting dashboard is live for everyone today. It should cut about 25% off your monthly reporting time, though the exact saving depends on how your team works.

Thirty-three words instead of 56. The 25% survived, along with "approximately" (now "about") and the workflow caveat — the two things a careless rewrite deletes for smoothness. Nothing new was invented.

That is the test: read the after, then scan the before for anything load-bearing that vanished. If you are rewriting model-generated drafts, the optimizer and the prompt generator handle the upstream half, and the mistakes article lists what people most often forget.