14 min · Power user
Structured Outputs and Quality Control
Demand a shape you can scan, then run a hostile second prompt that scores the draft against a rubric without adding facts or praise.
Learning objectives
- Specify a structured output with named headings, fixed columns, or tagged blocks.
- Run a hostile reviewer against a five-check rubric that cannot add facts.
- Separate the model's memo from the human decision, including a review date.
01
Why this matters
A wandering essay makes you hunt for the decision, the evidence, and the risk. A structured output puts those in places you can scan before you trust the prose. Quality control is a second prompt with a hostile job, not a request to polish the style. The reviewer who only flatters you has not reviewed anything.
02
Concept
Structured output — a shape you can scan, rather than a wandering essay — is a power-user requirement. Name the headings yourself, or demand a table with fixed columns, or use a tagged layout such as DECISION / EVIDENCE / RISK / NEXT. The shape does not make the content true. It makes a missing part obvious. If EVIDENCE is empty, you see it at once. A smooth paragraph can hide that emptiness for a page. Name that shape before the model writes.
Quality control — a second prompt whose only job is to find misses — should be hostile on purpose. Give it a rubric of five checks. Completeness: every section you named is present. Faithfulness to sources: each claim traces to a note you pasted. Specificity: names, dates, and numbers are concrete or marked unknown. Action: the next step has an owner. No new facts: the reviewer adds nothing the draft did not already contain. A pass on tone is not one of the checks.
The failure case is a reviewer that flatters the draft. It opens with praise, calls the work strong, and hides one soft suggestion at the end. That is not a review. Forbid praise in the prompt. Tell the reviewer to start with defects, to skip compliments, and to write none found if it cannot name a miss. If it still praises you, discard the review and rerun it. Praise feels helpful. It delays the repair.
AI-assisted decision making — the model prepares the memo and a human owns the choice — needs a line the draft cannot fake. The recommendation is input to a person, not a vote by the model. Name the decision owner. Set a review date, a real day when you will look again, and put that commitment somewhere you will see it. The NEXT block states the action. It does not declare that the action already happened. If the stakes are high, the shape is a briefing, not permission.
03
Real-world example
A vendor choice you can scan in one minute
A facilities lead had to choose between two service vendors and kept receiving essays that buried the price gap. She required this shape: DECISION, EVIDENCE, RISK, NEXT. DECISION held only the proposed vendor and the line that a named manager still had to accept it. EVIDENCE held three rows, each with a claim and the note it came from. RISK held the missing insurance certificate, marked unknown rather than assumed. NEXT named the manager, the question to ask legal, and a review date two weeks out. She then ran a hostile reviewer with the five checks: completeness, faithfulness to sources, specificity, action, and no new facts. The reviewer flagged a sentence that praised response times the notes never mentioned. She deleted that sentence. The manager could read the page without hunting.
04
Gemini demonstration
A realistic exchange. Open Gemini beside this page and try a version with your own facts.
You
Here is a draft of my vendor note. It says both vendors are excellent and I think we should proceed with Northline because they feel more modern. Review it kindly and improve the style.
Gemini
I am not going to praise this draft or polish the style. A kind review would flatter you and hide the miss. Structured output — a shape you can scan — shows the problem at once. DECISION: no decision is stated, only a feeling that Northline is more modern. EVIDENCE: there are no pasted notes, prices, or dates, so faithfulness fails. RISK: modern is undefined, and no risk is named. NEXT: there is no owner and no review date. Defects against the five checks: completeness fails, faithfulness fails, specificity fails, action fails, and I will not add facts to fill those holes. Rewrite the note in those four tags using only notes you paste. I will review again with no compliments. If I cannot find a miss after that, I will say none found.
You
DECISION: Propose Northline, pending approval by Operations Lead. EVIDENCE: Quote sheet shows Northline at 12 and Harbor at 14 per visit. RISK: Northline certificate of insurance is not in the file. NEXT: Operations Lead asks legal by Friday. Review date is 20 October. Run the hostile rubric. Do not add facts. No praise.
Gemini
Defects only. Completeness: the four tags are present, so that check passes. Faithfulness to sources: the prices match the quote sheet you just stated, and I am not adding a third vendor or a reason they differ. Specificity: the certificate is correctly marked missing, but Friday needs a calendar date so the action cannot slide. Action: the owner is Operations Lead, yet the review date of 20 October is not tied to a named checker. No new facts: I found no sentence that invents a testimonial or a service level. I am not calling the draft strong. Repair the two date lines, then stop. The decision still belongs to Operations Lead, not to this review.
05
Prompt breakdown
The prompt
Review the draft against this rubric only: completeness, faithfulness to sources, specificity, action, and no new facts. Start with defects. Do not praise. Do not add facts, numbers, or sources. Quote the failing line and name the failed check. If you find no miss, say none found.
- The rubric is closed, so the reviewer cannot wander into a new essay or a new recommendation.
- Starting with defects forbids the flattery failure, where praise hides a soft suggestion.
- The ban on new facts keeps the reviewer from becoming a second author.
- Quoting the failing line makes the repair local, instead of a vague request to be more specific.
06
Common mistakes
Usually
Please make this better and more professional.
Better
Return DECISION, EVIDENCE, RISK, and NEXT, and leave unknown where the notes are silent.
Usually
Check my draft and tell me it is ready for the director.
Better
List defects against the five checks, with no praise, and do not add facts.
Usually
The model chose Vendor A, so we are decided.
Better
The memo proposes Vendor A. A named person decides, and a review date is on the calendar.
07
Practice exercise
Put one real note into four tags
Done when: The page shows four tags, a named decision owner, a review date, and a review that starts with defects or says none found.
08
Challenge
Catch a flattering review
Paste a draft you care about and first ask for an encouraging review, so you can see the failure mode. Save that reply. Then run the hostile rubric: completeness, faithfulness to sources, specificity, action, and no new facts, with praise forbidden and no new facts allowed. Compare the two replies. Write three lines on what the encouraging review hid, what the hostile review caught, and which line you changed as a result. Do not let the second prompt introduce a fact the draft and your notes did not already contain.
Hint: If the hostile pass still opens with a compliment, reply with one line: delete the praise and start again from the first defect.
09
Key takeaways
- A named shape beats a wandering essay because gaps show up immediately.
- The five checks are completeness, faithfulness, specificity, action, and no new facts.
- Forbid praise, or the reviewer will flatter the draft and soften the miss.
- The model prepares the memo. A named human owns the decision and the review date.
10