奇创岛

From «Thinking, Fast and Slow»

Build a 'System 2 Agent' AI Toolkit

You'll design an AI prompt toolkit based on 5–8 core biases from *Thinking, Fast and Slow* (anchoring, availability, loss aversion, framing, overconfidence, etc.), so the AI acts as a 'slow thinking trigger' that questions your intuitive judgments when you make important decisions.

Final work

A 'System 2 Agent AI Toolkit' (5–8 bias-specific AI prompts + system prompt for each + usage scenario descriptions + boundary notes)

Estimated time

1.5–2 hr

Submitted

…

Your final work

Purpose:Before making an important decision, call on the AI to act as an 'external perspective' and 'slow thinking agent' through a preset bias identification system prompt, delaying System 1's intuitive impulses and performing a debiasing check.

Parts:

  • Toolkit usage guide (when to use / how to start / how many minutes to complete)
  • Bias identification checklist system prompt (general version, suitable for quick bias scan)
  • Consumer decision debiasing prompt (with system prompt + user input template + output format)
  • Investment judgment AI review prompt (with system prompt + user input template + output format)
  • Hiring / performance evaluation debiasing prompt (with system prompt + user input template + output format)
  • Job offer / career change evaluation prompt (with system prompt + user input template + output format)
  • At least 1 additional self-chosen scenario prompt (e.g., project initiation / parenting decision / entrepreneurial judgment)
  • Toolkit usage boundary notes (AI cannot replace decision-making itself / applicable and non-applicable scenarios)

Use cases:

  • · Before a large purchase, open the corresponding prompt, paste your decision description, and get bias-checking questions.
  • · Before adding to or cutting an investment position, use the AI to review whether your current judgment is dominated by loss aversion.
  • · After a job interview but before deciding to hire, use the AI to identify the halo effect and availability bias.
  • · When facing a career change or job offer decision, ask the AI to provide an external perspective and a pre-mortem.

Pick a topic

Pick the topic closest to you, or write a custom one when you submit.

Personal Life

Family / Parenting

Work / Project

Tools you'll use from the book

Bias Identification Checklist System Prompt

Condenses 5–8 core biases from *Thinking, Fast and Slow* (anchoring, availability, loss aversion, framing, overconfidence) into a system prompt that makes the AI act as a 'bias scanner,' reading your description and checking which biases might be interfering with your judgment.

How to use it here:

Serves as the 'general base prompt' for the toolkit. Can be used standalone or as the system prompt foundation for other specialized prompts. Usage: Paste this system prompt into the AI's system role area, then send your decision description to the AI and have it output a bias analysis report against the checklist.

Boundaries:

The AI's bias analysis is an auxiliary check, not a final verdict. Some intuitions are accumulated experience ('System 1 expert intuition'), not biases; the AI cannot distinguish this. You need to judge which bias diagnoses apply in your scenario.

Loss / Gain Perspective Parallel Prompt

Specifically designed to counter loss aversion and framing effects: forces the AI to restate your decision scenario from both a loss and a gain perspective, then points out which frame your original description leans toward.

How to use it here:

After describing your decision, send this prompt to the AI and ask it to write 3–5 points each for 'If I do X, I will gain...' and 'If I do X, I will lose...', then indicate which perspective you omitted in your description.

Boundaries:

This tool is not meant to make you pessimistic, but to fill in the half of the perspective you unconsciously omitted. If both perspectives are comprehensive, your original judgment may be correct.

Anchoring Detachment Prompt

Helps you identify and detach from anchors in your decision scenario: the first number, comparison object, or reference frame that appears. Then asks the AI to give an independent estimate from the perspective of 'assuming you never saw this anchor.'

How to use it here:

In decisions involving numbers or prices (consumption, salary negotiation, investment pricing), send your decision scenario and 'the first number you saw' to the AI. Let it identify whether that number constitutes an anchor and what the reasonable range would be after detaching from it.

Boundaries:

Anchoring detachment doesn't mean you should always reject the first offer—sometimes the first number is indeed reasonable. The goal is to consciously check before accepting it, not to follow it automatically.

Second Opinion Simulation Prompt

Makes the AI act as an 'external advisor who doesn't know you or your emotional background,' giving an independent evaluation of your decision description. It will raise questions you might not want to hear, rather than agreeing with your inclination.

How to use it here:

When you already have a preferred conclusion (e.g., 'I've basically decided to take this offer'), send your list of reasons to the AI and use this prompt to have it play the role of a challenger, specifically outputting '3–5 questions that someone who genuinely cares about you but disagrees with this decision would ask.'

Boundaries:

The second opinion simulated by the AI is deliberately biased toward criticism. It doesn't mean the AI thinks your decision is definitely wrong—you need to judge for yourself whether these challenges are valid and which ones deserve a serious response.

Assume I Might Be Wrong Prompt (Pre-Mortem)

A variation of the 'pre-mortem' recommended by Kahneman in the book: asks the AI to assume your decision turns out to be wrong in one year, then deduce the 3 most likely reasons for failure—forcing you to confront the blind spots you least want to think about before making the decision.

How to use it here:

Send your decision and reasons to the AI, using this prompt to have it output: 'Assuming this decision turns out to be wrong in 12 months, what are the 3 most likely reasons? Among these 3, which one can you verify right now?'

Boundaries:

The pre-mortem is not meant to make you negative, but to force you to check overconfidence. If you can refute all the failure reasons the AI gives, it means your decision passes the stress test.

Work rules

Your work MUST include

  • Must include a toolkit usage guide (when to open / how to start / how many minutes to complete one check).
  • Each prompt must include the full system prompt text (not a description like 'let AI identify biases' but a copy-paste-ready instruction paragraph).
  • Each prompt must include a user input template (telling the user where to fill in their decision description).
  • Each prompt must include the expected output format (how the AI should structure its response for quick reading).
  • The toolkit must cover at least 5 real usage scenarios, corresponding to different bias concepts from *Thinking, Fast and Slow*.
  • Must include boundary notes (when the toolkit is effective / ineffective / should not be relied upon).

Your work CANNOT just be

  • Cannot just list bias definitions from the book without providing copy-paste-ready AI prompts.
  • Cannot use 'let AI make the decision for you' instead of 'let AI check your biases' — this toolkit aids thinking, not outsources decisions.
  • Cannot have only a general prompt without scenario-specific versions.
  • System prompts cannot be vague like 'help me check biases' — must include specific bias names and check angles.
  • Cannot omit boundary notes, especially cannot imply that AI analysis is more authoritative than the user's own judgment.

AI can help you here

Round 1: Help me determine which scenarios the toolkit should cover

When to use: You want to build the toolkit but aren't sure which decision scenarios to cover, or which scenarios are most valuable to you.

I'm using '{{book title}}' to complete the '{{route name}}' project. The goal is to build an AI prompt toolkit that acts as a 'System 2 agent' when I make important decisions, questioning my intuitive judgments.

Based on my situation, please help me pick 3–5 decision scenarios most worth designing a dedicated prompt for.

My situation (please fill in):
- Which type of decision has made me most anxious or regretful in the last three months: [fill in]
- My life / work background (occupation, common decision types): [fill in]
- Which bias am I most vulnerable to (anchoring / loss aversion / availability / framing / overconfidence): [fill in, leave blank if unsure]

Optional scenarios (provided on the page):
[paste the list of topics from the page]

Please output:
1. The 3–5 most recommended scenarios and reasons
2. The 1–2 biases from *Thinking, Fast and Slow* most likely triggered in each scenario
3. Whether there is a priority order among scenarios (from easiest to design to most challenging)

Yellow placeholders need you to fill in before using the AI.

AI can help you organize ideas, but cannot make final judgments for you. Don't let AI fabricate experiences, cases, or misleading content.

Round 2: Help me design the three-part structure for scenario-specific prompts

When to use: You've chosen the scenarios to cover but don't know how to turn the book's bias concepts into actionable system prompts + user input templates + output formats.

My project is '{{route name}}' from '{{book title}}'.

My topic is:
{{topic}}

I've decided on the scenarios to design. Now I need help designing the three-part structure for each scenario prompt:
① System prompt (AI's role and task)
② User input template (what key information the user should fill in)
③ Expected output format (how the AI should structure its response for quick reading)

Please give a complete design example for the following scenario:
[fill in the scenario you want to design first, e.g., consumer decision debiasing / investment judgment review / offer evaluation]

Design requirements:
1. The system prompt must include specific bias names (e.g., anchoring effect, loss aversion) and the AI's check angle; don't use vague instructions like 'help me identify biases.'
2. The user input template should make it obvious what to fill in and how much, so I'm not staring at a blank box.
3. The output format should be structured (bullet points, labels) so I can read the AI's analysis in under 2 minutes.
4. The entire three-part prompt should be no more than 400 words, otherwise it's too cumbersome for actual use.

Yellow placeholders need you to fill in before using the AI.

AI can help you organize ideas, but cannot make final judgments for you. Don't let AI fabricate experiences, cases, or misleading content.

Round 3: Help me check the toolkit's completeness and usability

When to use: You've finished the first draft of the toolkit and want to confirm before submitting that the prompts are ready to use, the boundaries are clear, and there are no obvious gaps.

I'm submitting a project work on Shufang Island.

Book: '{{book title}}'
Project route: {{route name}}
My topic: {{topic}}

My work draft:
{{work draft}}

Please check against the following criteria:
1. Does the toolkit usage guide clearly state the trigger conditions and startup steps? (Not just vague phrases like 'use when making important decisions.')
2. Does the general bias identification system prompt include specific bias names and check logic? (Can it be copy-pasted and used directly?)
3. Does each scenario-specific prompt have a complete three-part structure? (System prompt + user input template + expected output format)
4. Is at least one prompt accompanied by a real usage record? (Test scenario + AI analysis + whether it changed the judgment)
5. Do the boundary notes cover 'non-applicable situations' and 'potential misuse'?
6. Can someone who hasn't read the book roughly understand how to use the toolkit?
7. Is it ready to submit?

Please output:
- Overall evaluation
- What's already good
- What must be changed
- What could be enhanced
- Suggested structure for the revised work

Yellow placeholders need you to fill in before using the AI.

AI can help you organize ideas, but cannot make final judgments for you. Don't let AI fabricate experiences, cases, or misleading content.

Submit my work