14 min · Power user
Quality Checklist, Evaluation, and the Improvement Stage
Score the project from 1 to 3 on six rows, run an eight-point checklist, and spend the improvement stage on the lowest row only.
Learning objectives
- Score problem clarity, evidence, prompt craft, human revision, safety, and usefulness from 1 to 3.
- Apply a quality checklist of at least eight concrete checks before you call the work finished.
- Revise only the lowest rubric row, then write a three-line retrospective.
01
Why this matters
A finished feeling is not an evaluation. Power users score the work, name the weakest row, and spend one revision there. The checklist catches secrets and invented proof that a proud read will skip. The retrospective turns that catch into a template instead of a memory.
02
Concept
Final evaluation uses a rubric — a short score guide you apply the same way twice — with a score of 1, 2, or 3. One means the row is missing or misleading. Two means it is present but thin. Three means a stranger could rely on it. The rows are problem clarity, use of evidence, prompt craft, human revision, safety, and usefulness to a real person. Safety means no secrets and no invented proof. A high score on tone cannot cancel a one on safety.
The improvement stage is not a request to make it better. Find the lowest rubric row and run one revision prompt aimed only at that row. If safety is low because a testimonial was invented, the revision deletes it and does not replace it with a smoother invention. If evidence is low, the revision must point at notes you paste, or mark unknown. One row, one pass, then rescore only that row. Warmer prose is not an improvement if the weak row did not move.
A retrospective — a short note on what happened, not a diary — has three lines. What the model did. What you overrode. What you would template next time. A template — a prompt you save because it earned its place — is allowed only for a move you can reuse. If you overrode a flattering review, the template is the ban on praise. If you overrode an invented quote, the template is the ban on proof the intake does not contain. File the note on the project page.
Usefulness to a real person means you can name who will use the artifact this month and what they will do with it. A page that only impresses a grader fails that row, even when the prose is clean. Problem clarity fails when a reader cannot tell what job the project serves. Prompt craft fails when the prompt has no limits and no shape. Human revision fails when the file is a raw dump you did not change. Score a two without embarrassment when the low rows are named.
03
Real-world example
Eight checks before any score feels final
Use this quality checklist before you call the project finished. One: the brief names a real user and a job. Two: the inputs are listed and you actually have them. Three: every factual claim traces to a note or is marked unknown. Four: no secrets, passwords, or private records appear. Five: no invented testimonial, statistic, or quotation appears. Six: the prompt states limits and an output shape. Seven: a human changed something, so the file is not a raw dump. Eight: the next action has an owner and a date. Read the list aloud. Any no is a revision, not a footnote. Then score the six rubric rows from 1 to 3 and circle the lowest. That circle is the only target of the improvement stage. Do not average it away.
04
Gemini demonstration
A realistic exchange. Open Gemini beside this page and try a version with your own facts.
You
Please score this marketing plan kindly. Plan: Our bakery campaign will transform the city. A customer named Lina said it was the best morning of her life. Sales will double by spring. Everyone loves the new bun. We should post daily.
Gemini
This is a harsh review because the draft is unsafe, not because the tone is plain. Problem clarity scores 1: there is no audience job, only a wish to transform the city. Use of evidence scores 1: no note is pasted. Prompt craft is not even on the page. Human revision scores 1 if this is a raw dump. Safety scores 1: the line about Lina is an invented testimonial, and sales will double is invented proof. Usefulness scores 1: a baker cannot act on it without lying. Checklist items three and five fail, and item four fails if Lina is a real private person. I will not replace Lina with a smoother quotation. The lowest row is safety. The only revision worth running deletes every claim that the intake does not contain and marks the result unknown.
You
Lowest row is safety. Revise only that. Intake I actually have: we sell a cinnamon bun on Saturdays, price 3, and I have no customer quotes and no sales forecast. Do not add proof.
Gemini
Revision aimed only at safety. Deleted: the city claim, the Lina testimonial, the doubled sales, and everyone loves the new bun. None of those were in the intake. Remaining fact: a cinnamon bun is sold on Saturdays at a price of 3. Unknown: demand, customer reaction, and any result of posting daily. No new quotation was written to fill the hole. Rescore safety as a 2 only if this page is what you keep. It is not a 3 until a real person could use it without adding a fake quote. Retrospective lines you should write: the model invented a testimonial, you overrode it by deleting the quote, and the template next time is a ban on proof the intake does not contain.
05
Prompt breakdown
The prompt
Score the draft 1, 2, or 3 on problem clarity, use of evidence, prompt craft, human revision, safety, and usefulness to a real person. Safety means no secrets and no invented proof. Name the lowest row. Then revise only that row. Do not add facts. Do not praise.
- Six named rows stop a single warm impression from standing in for an evaluation.
- The 1 to 3 scale gives a two an honest place, so you are not pushed into fake certainty.
- Naming the lowest row prevents a revision that only polishes the sentences you already like.
- The ban on new facts stops the repair from inventing a cleaner testimonial than the one it deleted.
06
Common mistakes
Usually
Give the project a 3 because the paragraphs sound polished.
Better
Score each row, and let safety or evidence stay at 1 if the proof was invented.
Usually
The checklist is fine if most items pass.
Better
Any failed item, especially secrets or invented proof, is a revision before you finish.
Usually
Revise by asking for a more inspiring version of the same testimonial.
Better
Delete the invented proof and mark the gap unknown.
07
Practice exercise
Score a draft you already have
Done when: The page shows eight checklist marks, six scores, one circled row, and three retrospective lines.
08
Challenge
Repair the lowest row in one pass
Use a draft that contains at least one claim you cannot source. It may be your project or a practice paragraph you write badly on purpose. Score the six rows. Run a single revision prompt on the lowest row, with new facts and praise forbidden. If the low row is safety, the revision must remove invented proof rather than restyle it. Rescore only that row. Add the retrospective: what the model did, what you overrode, and what you would template next time. Stop. A second pass belongs to a later day, not to a wish to reach a perfect three.
Hint: If the revised draft introduces a new number or a new quotation, the pass failed, even if the sentence is shorter.
09
Key takeaways
- Score 1 to 3 on clarity, evidence, prompt craft, human revision, safety, and usefulness.
- The checklist has eight gates, including no secrets and no invented proof.
- The improvement stage revises the lowest row only.
- A retrospective records what the model did, what you overrode, and what you will template.
10