Prompts That Work

The de-hype prompt for spec sheets and launch posts

Paste in a product page or press release and get back what is actually verifiable, what is weasel-worded, and what the page is carefully not telling you.

Use case
Research & buying decisions
Works with
Claude, GPT-class, local Qwen3 14B+
▶ The prompt
Below is marketing copy for a product. Strip the marketing.

Return four lists:

1. VERIFIABLE: every claim that could be tested or measured by a buyer,
   with the number if one is given.
2. WEASEL: every claim that sounds like a spec but cannot be tested as
   written. Flag "up to", comparisons with no named baseline ("2x faster"),
   and any use of "AI-powered" that does not say what the AI does.
3. MISSING: the specs a careful buyer would want that this page avoids
   stating.
4. QUESTIONS: the five things I should get answered before spending money.

Do not soften anything, do not add your own praise for the product, and do
not speculate about specs that are not on the page. MISSING means missing.

{paste the product page or press release}

Printer launch pages are where I use this most. A machine gets announced, the page says “up to 600mm/s” and “next-generation motion system,” and the one number I actually want (what does it weigh, what is the real chamber temp, is the bed mesh per-print or stored) is nowhere. This prompt makes the absence visible.

Why it works

  • The WEASEL category has definitions, not vibes. “Up to,” missing baselines, and undefined “AI-powered” are patterns a model can match mechanically, so it catches them consistently instead of when it feels like it.
  • “MISSING means missing” is the guardrail. Without it, models helpfully fill in specs from their training data, and now you have a hallucinated spec sheet, which is worse than the marketing you started with.
  • Separating VERIFIABLE from everything else changes how you read the page. A launch post with two verifiable claims and fourteen weasel ones has told you what you need to know.

What good output looks like

Short lists. The tell is when the MISSING list is longer than the VERIFIABLE list: that is a page written to be quoted, not evaluated. The QUESTIONS list usually becomes my actual pre-order research, verbatim.

Where it works (and doesn’t)

Any competent model handles this, including local 14B-class models, because it is a reading task, not a knowledge task. The one failure mode to watch on smaller models: they sometimes grade a claim as verifiable because a number is present, even when the number has an “up to” in front of it. Skim the VERIFIABLE list for smuggled weasels.

Marketing copy is written by people who know exactly which specs they are omitting. This prompt just reads it the way they wrote it.