Safe prompt rewriting checklist
Quick answer
Shortening a prompt is only safe when the task, the output and every limit survive. Check names, numbers and exact text, keep every “not” and “only after”, remove secrets before you paste, and compare the result with the original. Typo does the mechanical part of this check on every rule it applies.
What must not change
The task, the kind of output, the audience and the limits. A rewrite must not add requirements you did not ask for, and it must not remove the restrictions that decide whether a result is acceptable.
Names, numbers and exact text
Dates, amounts, percentages, links, e-mail addresses, names, code and quoted material often have exact meanings. If the AI needs a literal string to diagnose an error, or a citation as evidence, paraphrasing it does harm.
Negations and conditions
“Do not delete data” is not the same as “avoid deleting data if possible”. Conditions such as “only after approval” are not decoration either.
Private information
Never paste passwords, access keys or personal data into a tool you do not trust. Remove secrets before you shorten anything, and check what you copy into an AI tool.
Compare the result with the original
- Does it still ask for the same thing?
- Is every figure, date and “do not” still there?
- Did any code, link or identifier change?
- Were new assumptions added?
- Are the requested sources, tests and format still there?
- Does it still make sense to the reader?
When no rewrite is better
Prompts that are already direct, legal text, code and evidence-heavy prompts are often best left alone. Typo is allowed to return the original, and it does so whenever a check fails or the estimate does not go down.
How Typo applies this checklist
Protected text is masked before any rule runs, negations and strength words are counted before and after, and a rule that removes too much is rejected. What Typo cannot check is whether a model answers the same way, so read the result at levels 8 to 10.