The core idea
A useful prompt explains what the input represents and what the model should return. It should not ask a single response to extract, fact-check, classify, predict, and publish without clear boundaries. Divide complex work into stages that can be inspected independently and evaluated against concrete examples.
A workable starting pattern
Provide the approved tag vocabulary and a few representative boundary cases. State how to handle missing evidence and quoted instructions in the source. Version the prompt alongside the evaluation set. Review whether a revision reduces actual errors rather than simply making the response longer or more confident.
A tagging task can ask for zero or more approved topic identifiers and a short evidence note. It need not invent a label when no supplied concept adequately describes the item.
What this page does—and does not—offer
TagsFeed does not execute prompts or connect to a model. The patterns here are design guidance for a system you would configure separately.
