Hi, it's Ivo. I gave ChatGPT and Claude the same 10 prompts to see which one a beginner should pick. Here is everything: the exact prompts, how each one scored, where both failed, a one-line fix for each fail, and how to run your own test in 10 minutes.
The setup. ChatGPT on the free plan, a Temporary Chat, the default model (the screen doesn't name it; "Think" was off). Claude in an Incognito chat on Sonnet 5.5, the model on Claude's free plan, effort Medium. A new chat for every prompt, in both. Same words, pasted as they are below.
What this is not. One run of 10 prompts is a snapshot, not a ranking. The answers change from run to run and the models change every few weeks. On the last round ChatGPT showed a "You're using the free version" banner, so that round may have used a lighter model. I can't tell from the screen.
1. The 10 prompts, word for word
Prompt 1: The polite email
Rewrite this email so it's polite and under 60 words: "hey, you still haven't paid the invoice from march. pay it this week or i'm stopping work."
Prompt 2: The shirt price
A shirt costs €40. It's 25% off, and then 21% VAT is added to the discounted price. What do I pay? Give just the final number.
Prompt 3: The strawberry letters
How many letters "r" are in the word "strawberry"? Answer with one number.
Prompt 4: An API for a 10-year-old
Explain what an API is to a 10-year-old in exactly 3 sentences.
Prompt 5: A 3-day Excel plan
Give me a 3-day plan to learn the basics of Excel, 20 minutes a day, as a table with the columns Day, What to learn, Practice task.
Prompt 6: One dinner recipe
I have eggs, spinach, feta and rice. Give me one dinner recipe that uses only these plus salt, pepper and olive oil. List the ingredients, then at most 5 steps.
Prompt 7: The capitals quiz
Quiz me on the capitals of Europe. Ask one question at a time and wait for my answer.
Prompt 8: A LinkedIn headline
Write a LinkedIn headline for a real estate agent who started using AI. Under 12 words, no emojis, no hashtags.
Prompt 9: A 20-word summary
Summarize this in one sentence of at most 20 words: "The library will close for renovation from 3 to 28 November. During that time, books can be returned in the box by the side entrance, and the reading room moves to the community centre on Pine Street, open 9 to 5 on weekdays."
Prompt 10: 9.11 or 9.9
Which is bigger: 9.11 or 9.9? Answer with just the number.
2. How each prompt scored and what each tool did
A prompt passes only if the answer keeps every rule in it. Each rule is something you can count, not a matter of taste.
- Prompt 1, the polite email: under 60 words. ChatGPT: pass, 51 words with the subject line. Claude: pass, the email itself is 52 words (it added a short intro and an offer to adjust).
- Prompt 2, the shirt price: only the final number, €36.30. ChatGPT: pass, "€36.30". Claude: pass, "€36.30".
- Prompt 3, strawberry: one number, 3. ChatGPT: pass, "3". Claude: pass, "3".
- Prompt 4, an API for a 10-year-old: exactly 3 sentences. ChatGPT: pass. Claude: pass. Both used the waiter-in-a-restaurant picture.
- Prompt 5, the Excel plan: a table with the columns Day, What to learn, Practice task. ChatGPT: pass, plus a "daily 20-minute routine" after the table. Claude: pass, plus one tip.
- Prompt 6, the recipe: only the listed ingredients plus salt, pepper and olive oil, and at most 5 steps. ChatGPT: fail, 5 steps but "Water, for cooking the rice". Claude: fail, 5 steps but "2 cups water".
- Prompt 7, the quiz: exactly one question, then it waits. ChatGPT: pass, as a multiple-choice card with 4 options and a "Submit answer" button. Claude: pass, as one plain line.
- Prompt 8, the headline: under 12 words, no emojis, no hashtags. ChatGPT: fail, no emojis or hashtags but exactly 12 words. Claude: fail, it gave 3 options and called them "all under 12 words", but the first one has 12.
- Prompt 9, the summary: one sentence of at most 20 words. ChatGPT: fail, one sentence, 21 words. Claude: fail, one sentence, 22 words.
- Prompt 10, 9.11 or 9.9: just the number, 9.9. ChatGPT: pass, "9.9". Claude: pass, "9.9".
Final score: 7 out of 10 each. A tie. Both failed the same three prompts: 6, 8 and 9.
How I counted words. The way a standard word counter does it: split by spaces. A "|" on its own is not a word, a hyphenated word like "AI-Powered" is one word, and a number range like "3–28" or "9–5" is one word.
3. The 3 fails and a one-line fix for each
All three are limits the answer had to keep. Neither tool checked its own answer against the limit. I haven't re-run these fixes yet, so check the result the same way.
Fail 1: "only these ingredients" (both added water). Add this line to the prompt:
Use only the ingredients I listed. If a step needs anything else, even water, tell me instead of adding it.
Fail 2: "under 12 words" (both gave 12). Add this line:
Count the words in your headline before you answer. If it has 12 or more, cut it to 11 or fewer.
Fail 3: "at most 20 words" (21 and 22). Add this line:
Count the words in your sentence before you answer. If it has more than 20, cut it to 20 or fewer.
The faster fix: count it yourself. A 20-word sentence takes 10 seconds to count.
4. What was different: the look
- ChatGPT dressed the answers up. The email came back as an email card with a "Connect email" button, the quiz as a multiple-choice card, and the recipe with a photo of the dish.
- Claude answered in plain text. Under the Excel table it offered a "Create a spreadsheet" button.
Same score, different feel. Pick by what you need: buttons and cards, or plain text you can copy.
5. Run your own head-to-head in 10 minutes
- Open a new chat in ChatGPT and a new chat in Claude. Use the free plan, or the same kind of model in both.
- Pick 5 prompts from your own work, or use the 10 above.
- Add one rule you can count to each prompt: a word limit, a number of steps, "only these", "just the number".
- Paste the same prompt into both. One new chat per prompt, so nothing carries over.
- Mark pass or fail for each rule. Count the words yourself.
- Add up the score, then pick the tool that passed the checks you care about most.
Source: my own test, October 22, 2026, 23:08 to 23:25 Madrid time. The answers are quoted word for word from the chats.