Meta-Prompting and Advanced Techniques: AI Improves Your Prompts, Part 2

Hard decisions feel less stuck when you see them as branches. Ask AI: 'Map out my options as a decision tree with consequences.' Suddenly you can see the shape of the choice.

Some examples

• Prompt: 'I'm choosing between AP Bio and AP Stats. Build a decision tree.'
• AI surfaces consequences you hadn't considered.
• AI flags reversible vs irreversible branches.
• Decision trees beat pro/con lists for multi-step choices.

Try it!

Pick a real choice you're facing. Have AI build a decision tree 3 levels deep. Notice the new options.

A reverse outline is when you outline an essay AFTER writing it. Paste your draft into AI and ask 'outline what I actually wrote, one bullet per paragraph.' Weak structure becomes obvious.

Some examples

• AI shows you that paragraphs 3 and 5 say the same thing.
• AI exposes a paragraph with no clear point.
• AI reveals you forgot to address the counterargument.
• Reverse outlines beat asking 'is this good?'

Try it!

Take any essay or post you wrote. Have AI reverse-outline it. Find one weak paragraph and rewrite it.

Before you defend an idea in class, debate, or a pitch, have AI attack it as a smart skeptic. 'Find the strongest 3 objections to my plan.' You'll either patch them or change your mind.

Some examples

• Prompt: 'Steelman the opposite view of my essay.'
• Prompt: 'You are a hostile college admissions officer — pick apart this draft.'
• Prompt: 'List the 5 biggest holes in my business idea.'
• Devil's advocate prompts beat asking friends who'll be nice.

Try it!

Pick an opinion you hold. Ask AI for the 3 strongest counterarguments. Try to rebut each one.

Models will happily make things up with the same tone they use for facts. Asking for explicit confidence levels — or 'flag anything you're unsure about' — gets the model to mark the risky bits, which you can then verify.

Some examples

• 'Answer this trivia question and rate your confidence 0-100.' Below 70 = double-check.
• 'Write this code and flag any function calls you're not 100% sure exist.'
• 'List 5 historical facts and mark each as confirmed or 'I might be misremembering.''
• 'Translate this paragraph and flag any phrases where you guessed at meaning.'

Try it!

Ask Claude or ChatGPT a tricky factual question. Add 'rate your confidence 0-100 and explain.' Notice what it flags.

self-critique catches around half of small errors for free

Some examples

• Generate, then critique, then revise
• Ask for a confidence score per claim
• Reject if confidence is low

Try it!

Open your favorite AI tool and try one of the examples above. Pick the one that matches what you are actually working on this week. Spend 10 minutes, no more. Notice what worked and what did not — that's the real lesson.

Self-consistency = sample multiple answers, pick the most common. Cheap accuracy boost on reasoning tasks.

Some examples

• Run with temperature 0.7 so you get variety.
• Run 5–10 times.
• Take the mode (most common answer).

Try it!

Pick a logic puzzle. Run it 5 times. See if the majority answer is right.

Understanding "Self-consistency: ask the same question 5 times" in practice: Prompting is a skill: the more specific and structured your input, the more useful the output. For tricky problems, run the prompt multiple times and take the majority answer — accuracy jumps — and knowing how to apply this gives you a concrete advantage.

• Use role, context, task, and format in every prompt
• Iterate: treat first outputs as drafts, not finals
• Use few-shot examples for complex formatting tasks
• Test prompts at different temperatures for creative vs. factual tasks

1. 1Rewrite one of your best prompts using role + context + task + format
2. 2Ask an AI to critique your prompt and suggest improvements
3. 3Compare outputs from two models using the same prompt

Meta-prompting = asking the model to draft, critique, or improve a prompt before you use it.

Some examples

• Describe your goal and audience.
• Ask 'write the best prompt for this.'
• Test, then ask AI to critique its own prompt and try v2.

Try it!

Pick your most-used prompt. Have AI write a v2. A/B test.

Understanding "Meta-prompting: ask AI to write your prompt" in practice: Prompting is a skill: the more specific and structured your input, the more useful the output. AI is great at writing better AI prompts than you are. Use it — and knowing how to apply this gives you a concrete advantage.

• Use role, context, task, and format in every prompt
• Iterate: treat first outputs as drafts, not finals
• Use few-shot examples for complex formatting tasks
• Test prompts at different temperatures for creative vs. factual tasks

1. 1Rewrite one of your best prompts using role + context + task + format
2. 2Ask an AI to critique your prompt and suggest improvements
3. 3Compare outputs from two models using the same prompt