How to Make a Decision with AI Using Five Criteria
A five-criteria skill for a real decision with AI. You name the options and the scale, ChatGPT, Claude, or Gemini fills a grid, UNKNOWN stays empty, and you write the choice in one sentence.
Anyone facing a real fork can make a decision with AI using five criteria in 15 minutes: name the options, lock a scale, fill a grid, leave UNKNOWN empty, then write the choice yourself. Run it in ChatGPT, Claude, or Gemini before you ask “what should I do.”
“What should I do” is a horoscope prompt. It invites a confident stranger to live your constraints.
A decision is a score you can defend, plus a sentence you can say out loud. If you want the pattern for reusable loops, start with what a skill prompt actually is and then lock the five below.
Five criteria, then a grid, then your sentence
Do not describe your feelings in the first message and hope the model finds a structure. Feelings are input. They are not columns.
1. Name the decision in one line
Verb plus object plus deadline. “Choose a lease by Friday.” “Pick a contractor for the roof this month.” Not “figure out housing.”
Instruct: “Repeat the decision in my words. List no options yet. List no criteria yet.”
If the model offers options, restart. Unsolicited options are a winner essay in disguise. If the job is still a ramble, force a one-page brief first.
2. You write five criteria and the scale
Five. Each one a thing you can observe.
Good: cost this year, hours of your time, reversibility, who else is affected, one hard constraint already decided.
Bad: synergy, alignment, “feels right,” future-proof, vibe.
Send the five with a 1–5 scale and what 1 and 5 mean in a short phrase. Tell the model: “Do not add a sixth. Do not rename mine. If a criterion is vague, ask me one clarifying question and stop.”
One question. Not a workshop.
3. You name two or three options. Freeze the set.
You name them. “Option A stay. Option B move. Option C delay 90 days.”
“Score only these. If you think of a clever fourth, write PARKED and do not score it.”
Clever fourths are how “delay” becomes “move to a different city the model likes.”
4. Fill the grid. Ban advice.
“One row per option. One column per criterion. Each cell is a 1–5 plus a seven-word fact from what I provided, or UNKNOWN. No advice under the table. No ‘I recommend.’ No weighted total unless I already sent weights.”
UNKNOWN is the point. It shows the fact you still owe yourself.
If you have a document in the mix, ground the cells the way you study a PDF without fake quotes: if you did not paste the fact, the cell cannot claim it.
Spot-check the column you care about most. If a score uses a fact you did not give, mark the column FAIL and refill that column only.
5. Write the choice in your mouth
Read the grid out loud. Cross one option. Keep one. Write: “I choose B because of criteria 2 and 5. I accept UNKNOWN on 4.”
Send that sentence back: “Do not recast my choice as advice. Do not add a pep talk. Confirm the grid is unchanged.”
Then stop. A pep talk is how you outsource the feeling you still have to live with.
Save the five-criteria law, not this month’s options. The Promptcrates tools for reusable skills are the place.
Reusable skill card
Trigger: You are about to ask ChatGPT, Claude, or Gemini what you should do and you can already predict a wise paragraph.
Input: One decision sentence, five criteria with a 1–5 scale you defined, two or three options you named, optional weights you wrote first.
Output: A grid of scores plus seven-word facts or UNKNOWN. PARKED if a fourth option appears. No recommendation.
Stop: After you write the choice in one sentence. No recap. No “if I were you.” No extra option.
Takeaways
- Decision line first. Criteria second. Options third. Grid fourth. Your sentence last.
- Five criteria you can observe. Vibe is not a column.
- UNKNOWN is a to-do, not a 3.
- You choose. ChatGPT, Claude, and Gemini do not advise.
- Save the law as a skill. Options expire.
Frequently Asked Questions
Can I ask the model which option it would pick?
No. That question is how a grid becomes a guru. You write the choice. The model keeps the table still.
What if I only have two options?
Two is fine. Keep five criteria. A two-by-five grid still beats a vibes paragraph.
Should I weight the criteria?
Only if you write the weights before scoring. Do not let Claude or ChatGPT assign weights after it sees the options.
What if a criterion cannot be scored from what I know?
UNKNOWN. Then you either fetch the fact or accept the gap. Do not let the model fill it from “typical cases.”
Can I reuse last month’s criteria?
Reuse the loop, not the five words, unless the job is the same. A hiring grid is not a lease grid.