Enrichment That Says Unknown Instead of Guessing

Enrichment That Says Unknown Instead of Guessing

Makes extraction steps return NOT_FOUND plus the source sentence, so a filled column is never quietly invented.

airops data-enrichment scraping grounding hallucination validation

What it is for

Bulk enrichment fails silently: the column looks complete because the model filled the gaps from memory instead of the page. This forces an explicit NOT_FOUND, demands the source sentence beside every value, and stops extraction from running on a failed fetch. Use it for any table that sales, pricing or a published page will later rely on.

The skill

When a workflow fills columns from scraped pages or search results, the risk
is not an empty cell. It is a filled one that is wrong, because nothing
downstream can tell the difference.

Put this in the instruction of every extraction step:

  Use only the content provided below. Do not use anything you know about
  this company, person or product from elsewhere, even if you recognise it.
  For each field return either the value, or exactly: NOT_FOUND
  Return a value only if it appears in the provided content. A value you
  inferred, calculated, translated or remembered is NOT_FOUND.
  With every value, also return the sentence you took it from, copied
  exactly.

Then, in the sheet:
- Keep the source-sentence column. It is how you catch the model quoting a
  navigation menu, a cookie banner, a testimonial, or a different company
  mentioned halfway down the page.
- Count NOT_FOUND per field before using anything. A field that is ninety
  percent filled from bad sources is worse than one honestly forty percent
  filled, because it will be believed.
- Spot-check twenty rows against the live page, including three that came
  back NOT_FOUND, to see whether the value was genuinely absent.
- Never let a later step treat NOT_FOUND as an empty string and build a
  sentence around it.

For the fetch step:
- If the page failed, hit a login wall, returned a cookie interstitial, or
  came back under two hundred words, mark the row FETCH_FAILED and stop.
  Do not run extraction on nothing. Extraction on nothing is precisely where
  confident inventions come from.

Failure this prevents: a clean-looking enrichment that sales calls from or a
page publishes, where a tenth of the values were never on the page at all.

How to install it

1. Open the tool and start a new conversation, project or template.
2. Paste the text below wherever the tool keeps standing instructions
   — usually called system prompt, instructions, or persona.
3. Save it so it applies to every run rather than a single message.

Did it work for you?

Where it came from

Skills app Original