About This Article
This article is created using an automated generation workflow leveraging generative AI. It is organized as an experiment where the same prompt is sent to an LLM with only a single change to the output format, observing not whether the response is good or bad, but whether it results in a format that is easier for humans to verify.
Verification Status: 📘 Prompt design checked against primary specification general principles; model-specific output unverified.
When making requests to an LLM, in addition to giving detailed instructions on the content,fixing the output format makes it easier for humans to review.The key is not some magic trick to make the AI always correct, but rather transforming the output into a shape where omissions and ambiguities are easier to spot.
First, Test with the Same Dummy Text
The input text is fixed.
会議メモ: A社への見積は金曜まで。担当は田中さん。 価格表は最新版か未確認。部長承認が必要。
Initially, the request is made as follows.
この会議メモを整理してください。
Next, only theoutput formatis changed.
この会議メモを整理してください。 出力は次の4項目だけにしてください。 - 作業 - 担当 - 期限 - 未確認事項 不明な項目は「不明」と書いてください。
What to Look For
We are not comparing the fluency of the text. Instead, we check whether items like "Price list updated status unconfirmed" remain as unconfirmed items, and whether the AI hallucinates owners or deadlines for items where none exist.
Why It Becomes Easier to Verify
In free-form text, information gets buried within sentences. Fixing the fields allows human reviewers to verify whether all four columns have been checked. However, format specification does not guarantee factual accuracy; plausible errors can still occur even with the exact same format.
Changing Just One Place
Only the final line is changed as follows.
不明な項目は推測せず「要確認」と書いてください。
Compare "Unknown" versus "Action Required" to see which one makes it easier for downstream personnel to take action.
For Professional Use
For meeting minutes, primary inquiry triage, incident reports, and vendor questionnaires, fixing fields such as "Fact / Guess / Unconfirmed" or "Assignee / Deadline / Next Check" makes human review much easier to operationalize. Please check the AI service terms and your organization's policies regarding whether confidential or personal data can be inputted.
