Prompt specificity amplifier
The fastest way to get inconsistent output from a model is to use vague words. “Write a good summary,” “use an appropriate tone,” “include some examples” — each hands the decision back to the model, and it guesses differently every time. This tool scans your prompt for vague qualifiers, fuzzy quantifiers, and hedges, then suggests specific, measurable replacements so the model has a concrete target it can actually hit. You stay in control: it flags and recommends, you decide the exact criterion.
The consistency problem with vague instructions
Vague words are not just aesthetically unsatisfying — they are a reliability problem. When a prompt says “good,” the model fills in what “good” means based on whatever context is available. On different inputs, in different parts of a conversation, or across different model versions, the threshold it applies varies. The result is output that seems inconsistent when the real cause is an underspecified instruction.
This is also why vague prompts are hard to debug. If your prompt says “write a comprehensive analysis” and the outputs vary in depth, you cannot tell whether the problem is the word “comprehensive” (underspecified) or something else. Replace it with “write an analysis covering at least five distinct points with a minimum of 400 words” and variation in that dimension becomes clearly attributable.
The three categories of vague language
Subjective qualifiers — Words like “good,” “better,” “appropriate,” “relevant,” and “high-quality” require the model to apply a standard you never defined. Each one is a gap between what you want and what the model will try to produce.
Fuzzy quantifiers — Words like “some,” “few,” “several,” “many,” and “a number of” leave quantity undefined. “Include some examples” could mean two or ten.
Hedges — Phrases like “try to,” “if possible,” “where applicable,” and “as needed” signal that the instruction is optional. Models often treat hedged instructions as lower priority and skip them under pressure.
How it works
You paste a prompt and optionally name the domain. The tool tokenizes the text and matches against a built-in dictionary of vague terms grouped into the three classes above. For each hit it surfaces the offending word with context and a concrete replacement pattern. It also computes a specificity score so you can see progress as you tighten the wording. Everything runs locally in your browser.
Tips and examples
- Make it checkable. If you cannot verify whether the output met a word’s bar, replace the word. “High quality” → “no spelling errors and under 200 words.”
- Replace quantifiers with numbers. “Some examples” → “exactly 3 examples.”
- Kill the hedges. “Try to keep it short” → “keep it under 50 words.” Hedges invite the model to ignore the instruction.
- Label the domain. Noting that you are writing a legal or marketing prompt tags the copied report so you know which prompt it belongs to.