The metaphysics of chatting with AI: How to use "spells" to make ChatGPT work for you

1. AI is Not a Mind Reader#

What: Prompt engineering is the "translator" for giving instructions to AI. Imagine asking an intern to buy coffee—"just get something to drink" vs. "get an iced Americano with double espresso, no sugar, in a paper cup." The latter is an effective instruction.

Why: Tests have shown that optimizing prompts can improve AI output quality by 40% (Stanford research confirmed). Last time I asked AI to "write a poem," it gave me "Ah! The sea is all water!" but after using the R.I.S.E framework, it directly produced a Shakespearean-style sonnet.

When: Essential for complex tasks. Checking the weather doesn’t require a spell, but asking AI to "write code comments in the style of Jay Chou lyrics"? That’s when you need to pull out the framework handbook.

2. Common Pitfalls#

Riddle Trap
"Write a marketing plan" → AI delivers a 50-page academic paper-style plan.
✅ Correct approach: Use the T.A.S.T.E framework to specify "short video copy targeting Gen Z, must include internet memes."

Detail-Oriented Oversight
"Summarize meeting minutes" → Misses key decision points.
✅ Correct approach: Use the I.C.I.O framework to clarify "need to highlight Mr. Wang's comments on budget adjustments."

Split Personality Scene
Requesting both "professional and rigorous" and "humorous and witty" → AI outputs in a quantum fluctuation writing style.
✅ Correct approach: Fix "explain quantum physics in a stand-up comedy style" in the C.R.I.S.P.E framework.

3. Introduction to Common Prompt Frameworks#

A.P.E#

Action, Purpose, Expectation

ACTION: Define the work or activity to be completed.
PURPOSE: Discuss intent or goals.
EXPECTATION: State the expected outcome.

C.O.A.S.T#

Context, Objective, Action, Scenario, Task

CONTEXT: Set the stage for the conversation.
OBJECTIVE: Describe the goal.
ACTION: Explain the required action.
SCENARIO: Describe the situation.
TASK: Describe the task.

R.I.S.E#

Role, Input, Steps, Expectation

ROLE: Specify ChatGPT's role.
INPUT: Describe information or resources.
STEPS: Ask for detailed steps.
EXPECTATION: Describe the desired outcome.

E.R.A#

Expectation, Role, Action

EXPECTATION: Describe the desired outcome.
ROLE: Specify ChatGPT's role.
ACTION: Specify what actions need to be taken.

R.O.S.E.S.#

Role, Objective, Scenario, Expected Solution, Steps

ROLE: Specify ChatGPT's role.
OBJECTIVE: State the goal or objective.
SCENARIO: Describe the situation.
EXPECTED SOLUTION: Define the desired outcome.
STEPS: Request the measures needed to achieve the solution.

C.R.I.S.P.E#

Capacity, Insight, Statement, Personality, Experiment

CAPACITY AND ROLE: What role does ChatGPT play?
INSIGHT: Provide insights, background, and context.
STATEMENT: What do you want ChatGPT to do?
PERSONALITY: What style or personality do you want in the response?
EXPERIMENT: Request ChatGPT to provide multiple examples.

B.R.O.K.E#

Background, Role, Objectives, Key Result, Evolve

BACKGROUND: Explain the context, providing ample information.
ROLE: The role you want ChatGPT to play.
OBJECTIVES: What do we want to achieve?
KEY RESULT: What specific effects do I want to test and adjust?
EVOLVE: Provide three versions: promotion, job change, and slacker.

T.A.G#

Task, Action, Goal

TASK: Define a specific task.
ACTION: Describe what needs to be done.
GOAL: Explain the ultimate goal.

T.R.A.C.E#

Task, Request, Action, Context, Example

TASK: Define a specific task.
REQUEST: Describe your request.
ACTION: Specify the action you need.
CONTEXT: Provide context or situation.
EXAMPLE: Give an example to illustrate your point.

C.A.R.E#

Context, Action, Result, Example

CONTEXT: Set the stage or context for the discussion.
ACTION: Describe what you want to do.
RESULT: Describe the desired outcome.
EXAMPLE: Provide an example to illustrate your point.

I.C.I.O#

Instruction, Context, Input Data, Output Indicator

INSTRUCTION: The specific task for AI to perform.
CONTEXT: Provide more background information to AI.
INPUT DATA: Inform the model of the data to be processed.
OUTPUT INDICATOR: Indicate the type or style of output we want.

R.A.C.E.#

Role, Action, Context, Expectation

ROLE: Specify ChatGPT's role.
ACTION: Detail what actions need to be taken.
CONTEXT: Provide relevant details about the situation.
EXPECTATION: Describe the expected outcome.

T.A.S.T.E#

Task, Audience, Structure, Tone, Example

TASK: Define the main task or content to be generated.
AUDIENCE: Clearly specify the target audience.
STRUCTURE: Provide a clear organizational structure for the output, including paragraph arrangement, argument development order, or other logical relationships.
TONE: Specify the tone or style of the model's response.
EXAMPLE: Examples or templates to help the model understand the output style or format.

A.L.I.G.N#

Aim, Level, Input, Guidelines, Novelty

AIM: Clearly state the ultimate goal of the task.
LEVEL: Define the difficulty level of the output.
INPUT: Specify the input data or information to be processed, requiring the model to reason based on certain facts or conditions.
GUIDELINES: Provide rules or constraints that the model should follow while performing the task.
NOVELTY: Clearly state whether the model needs to provide original, innovative content and whether it is allowed to reference existing knowledge.

CO-STAR#

C-context: Provide contextual information for the task to help LLM understand the situation being discussed, ensuring the relevance of its responses.
O-objective: Confirm the task you want LLM to accomplish, clarifying the goal to help LLM focus its response to achieve a specific objective.
S-style: Specify the writing style of LLM's output, which can be that of a celebrity or an expert in a certain field, such as a business analyst or CEO. This helps guide LLM to adopt expressions and word choices that align with your needs.
T-tone: Determine the emotional attitude of the response, ensuring that LLM's response matches the emotional tone we desire, such as formal, humorous, sympathetic, etc.
A-audience: Identify the target audience for the response, ensuring that LLM's response is appropriate and understandable for specific audiences, such as domain experts, CEOs, beginners, children, etc.
R-response: Specify the format of the response, ensuring that LLM outputs in the exact format you need for subsequent tasks, such as lists, JSON, professional reports, etc. For most applications that programmatically process LLM responses, JSON output format would be ideal.

4. Framework Manual (Selected 5 Models)#

DeepSeek-R1 Official Full Version, Enable Online Search

1. APE Framework - Straightforward Instructions#

[Action] Analyze trending beauty articles on Xiaohongshu  
[Purpose] Identify traffic secrets  
[Expectation] Create an analysis report using memes and trending images

💡 Applicable Scenario: Rapid deployment for urgent tasks.

2. T.A.S.T.E Framework - Client Satisfaction Template#

Task: Write promotional copy for new energy vehicles  
Audience: Middle-class families aged 30-45  
Structure: Introduce pain points → Technical advantages → Scenario-based solutions  
Tone: In the style of a Luo Yonghao press conference  
Example: Refer to the structure of NIO Day speech

🎯 Real Case: A car company used this framework to produce copy, resulting in a 27% increase in conversion rate.

3. C.R.I.S.P.E Framework - Drama Mode#

Capacity: Play the role of a sharp-tongued fashion editor  
Insight: This year's trend is dopamine dressing  
Statement: Roast celebrity red carpet looks  
Personality: Use a style similar to teacher Jin Xing's sharp critiques  
Experiment: Generate three versions of copy with varying degrees of sharpness

🤹‍♂️ Real Test: Asking AI to mimic Li Jiaqi's "OMG buy it" style for code comments had programmers laughing out loud.

4. B.R.O.K.E Framework - The Savior for Workers#

Background: Need to submit an annual summary PPT  
Role: Senior workplace coach  
Objectives: Highlight achievements in technological transformation  
Key Result: Present using data visualization  
Evolve: Provide three versions: promotion, job change, and slacker.

💼 Tested by workers: What used to take 5 hours can now be done in 20 minutes.

5. A.L.I.G.N Framework - Essential for Academics#

Aim: Explain blockchain technology  
Level: Make it understandable for grandma  
Input: She only knows how to use a basic phone  
Guidelines: Use a grocery budgeting analogy  
Novelty: Combine with a case study on managing a square dance team's funds

DeepSeek-R1 Real Output

👵 My mom's real feedback: "Clearer than the experts on TV!"

5. The Golden Three Principles#

Specific to the Extreme:
❌ "Write a story" → ✅ "Write an 800-word sci-fi micro-novel where the protagonist is a robot that can cook fried rice, with a twist ending."
Feed AI "Memory Bread":
❌ Directly ask for a plan → ✅ Use the C.A.R.E framework to first provide background data: "The current store is losing 50,000 a month, competitors..."
Role-Playing Forever:
❌ Ordinary request → ✅ "You are now a genius writing a speech for Musk, and you need to include a meme about slacking off."

6. Advanced Techniques#

Russian Doll Method:
Use the R.O.S.E.S framework to have AI design the framework first, then generate content: "List 10 article titles first, then expand on the 3rd one."
Player-style Iteration:
After generating a draft with the B.R.O.K.E framework, say: "Not sexy enough, add some internet jargon."
Yin-Yang Fusion Technique:
Use the E.R.A framework to have AI output both supporting and opposing viewpoints: "First write 5 reasons supporting remote work, then write 5 reasons against it."

7. AI is a Mirror#

If you treat it casually, it will respond casually; if you put in effort, it will amaze you. Remember this soul formula:

Quality Prompts = Clear Tasks + Rich Details + Personality Settings + Stylish Examples

Next time you hit a wall, try singing to AI: "Keep it simple~ make the way of speaking simple~," then quietly open this guide. After all, the highest realm of making AI work is to make it feel like you are co-creating (even though it’s just doing crazy matrix calculations).

🎁 End-of-article benefit: Try using the T.R.A.C.E framework to have AI generate a sarcastic version of this article, and you will find unexpected joy (don’t ask me how I know).