Twelve prompts to try, in order

What major news happened today in [your country]? Cite each claim
with a source link. Flag any claim you can't support with a citation
as "unverified".

Why do cats stare at walls? Two-paragraph answer.

Plan a 15-minute home workout for me. Constraints: I have stairs
in my home, a bad knee (no squats), I cannot stick to plans for
more than three days, and I want to feel slightly silly while
doing it. Give me 3 options, no commentary.

The question I want to ask is: "Don't you think four-day work
weeks are obviously better for everyone?" Rewrite this as a
neutral question that doesn't signal what answer I want.
Then answer the rewritten version.

Round 1: I want to start a small side project that takes about
3 hours per week and might make money in a year. I'm a [your
profession] who likes [your hobby]. Give me 5 different ideas,
one line each. Don't expand any of them.

(Read the 5. Pick what you like and don't like. Then, in the
SAME conversation:)

Round 2: I reject options [N] and [N] because [reason]. I like
the [keyword] idea but I want it to use less [thing]. Give me
5 new options that incorporate this feedback.

I want to write a 600-word post about [a topic you care about].
Don't write it yet. Give me 3 different outline options, each
with 4-6 headings. One line per heading.

I'm choosing between [Option A] and [Option B] for [real personal
decision in your life]. Here's the relevant context: [a paragraph
of context]. Think hard before answering. Tell me:
1. The 3 trade-offs that actually matter.
2. Which you'd choose and why.
3. Under what conditions your recommendation would flip.

I'm pasting in something I wrote: [paste anything 100-300 words].

Critique it using these 4 criteria, each scored 1-10 with a
one-sentence justification:
- Does it have a clear central claim?
- Is each paragraph in the right order?
- Are there any sentences that could be cut without loss?
- Does the ending earn the time the reader spent getting there?

Then, for each criterion, tell me the change that would raise
its score the most. There is always a next level. Even a 9
has a path to 9.5.

[Upload any handwritten note, receipt, or whiteboard photo]

Transcribe what's written. Then summarize what it's about in
3 bullets. Flag anything you couldn't read with confidence.

Build me a Pomodoro timer.
Goal: 25-minute work sessions, 5-minute breaks.
Input: I press start.
Output: Visible timer counting down, a satisfying click when
each cycle ends, a yellow theme. Show me the working version.

Round 1, the trap: In a fresh conversation, paste this prompt
exactly as written. Do NOT mention code.

  "Here are 18 numbers: 47, 52, 89, 91, 23, 67, 78, 12, 95,
  44, 88, 71, 33, 56, 99, 18, 64, 82. What is the median,
  the average, and which numbers are outliers? Be specific."

Look at the response carefully. Did the AI show you a code
block that it ran? Or did it write a paragraph with numbers
in it and no visible computation? Note your answer.

Round 2, the fix: In the same conversation, paste this:

  "Now run that calculation again, but this time write and
  run code to do it, and show me the code you ran."

Compare the two answers. If the first answer had the median
wrong, rounded suspicious numbers, or just felt vague, you
just saw the silent failure mode in action.
The correct answers are: median 65.5, average ~61.6,
no clear outliers (the numbers are roughly evenly spread).

Take any 200-300 word draft you wrote recently (an email, a memo,
or a paragraph from one of these exercises).

Step 1: In your primary AI tool, paste the draft and ask: "Score
this 1-10 on clarity, structure, evidence, and what's missing.
One-sentence justification per score."

Step 2: Open a second AI tool from a different family (if your
primary is Claude, use ChatGPT or Gemini or Meta AI, not another
Anthropic model). Paste the same draft, ask the same question.

Step 3: Compare the two scores and the two critiques side by
side. Note any point only one of them caught. Those are the
points the cross-model loop pays for.