People Who Aren't Swayed by AI Ask This: 10 ChatGPT Prompt Hacks

People Who Aren't Swayed by AI Ask This: 10 ChatGPT Prompt Hacks

Introduction: Why Does ChatGPT Fail Even Though It's "Smart"?

ChatGPT takes the words you throw at it and returns a "seemingly optimal solution." While this is convenient, if you ask in a sloppy manner, you'll get a sloppy answer—this obvious fact is surprisingly often overlooked.


What the ZDNET article conveys is not so much a trick as it is "etiquette to prevent AI from misunderstanding." In other words, it's akin to the instruction techniques for delegating tasks to a human subordinate.


Looking at the reactions on social media, you see a mix of voices saying, "It's not that ChatGPT is bad; often it's the way we ask that's the problem," and "However, AI can confidently make mistakes." In other words, the discussion is about the need for a set of "techniques to improve output quality" and "techniques to prevent accidents."


1. Replace Abstract Words with "Concrete" Ones

Words like "nicely," "clearly," and "seemingly" offer too much freedom for AI.
Here are three tips for concretization:

  • Purpose: What is the text for? (e.g., internal proposal, social media post, job interview)

  • Audience: Who is it for? (e.g., beginners, decision-makers, field staff)

  • Expected Standard: What is considered "good"? (e.g., conclusion-first, bullet points, with evidence, including counterarguments)


Example:
× "Come up with a marketing strategy"
○ "Increase free trial registrations for B2B SaaS. Budget of 300,000 yen per month, duration of 6 weeks. Provide 5 plans with execution steps, KPIs, and potential risks."


2. Provide Assumptions and Background Information "First"

ChatGPT fills in the background on its own. If the fill-in is correct, it's great; if not, it's a disaster.
Therefore, fix the "worldview" first.

  • Current Situation (What is happening)

  • Constraints (Time, budget, rules, available means)

  • Undesired Outcomes (Examples of NG or pitfalls)


The more background you provide, the longer it gets, but the point is that it reduces rework and speeds up the process.


3. Specify the Output Format in Advance

Specifying the form, such as "in an article," "in a table," or "conclusion→reason→example," reduces variability.
The following specifications are particularly effective:

  • Heading Structure (Equivalent to h2/h3)

  • Number of Bullet Points (e.g., up to 3 per item)

  • Character Count (e.g., introduction 200 characters, each section 400 characters)

  • Tone (e.g., polite, assertive, suitable for social media)


Once the format is decided, AI can fill it in without hesitation.


4. Assign a Role (Persona) to Fix the Perspective

Assigning roles like "You are an editor," "You are an IT manager," or "You are a CFO" changes how information is gathered, word choice, and risk sensitivity.


The trick is to specify not just the "title" but also years of experience and achievements.

Example:
"You are a manager responsible for both lead acquisition and nurturing, having worked in B2B marketing at a SaaS company for 10 years."


5. Clearly State Constraints (Including What Not to Do)

Constraints do not lower quality; they actually improve it.
The reason is simple: without constraints, AI will produce the "broadest correct answer."

  • Permissible Media (X, note, press, YouTube, etc.)

  • Budget Limit

  • Prohibited Expressions (Exaggerated advertising, definitive statements, medical efficacy, etc.)

  • Legal and Compliance Notices


Including rules like "Do not include links in the text" here makes them less likely to be broken.


6. Show Good and Bad Examples to Share Evaluation Criteria

Preferences for writing vary from person to person. AI can only guess "your preference."
Therefore, provide examples, even short ones, to align the criteria.

  • Good Example (This tempo, this vocabulary, this density)

  • Bad Example (Dislike this phrasing, too abstract is NG)


This is similar to the practical instruction of "Here's a reference article."


7. Don't Finish in One Go. Break Down the Process to "Organize"

"First, the outline," "Next, key points for each heading," "Finally, the text"—this division is powerful.


When AI tries to create a finished product all at once, consistency may break down, or it may conveniently fill in gaps.


By breaking down the process, you create intermediate deliverables that can be checked, making corrections easier.


8. Let It Ask Questions First (Identify Missing Information)

This is subtly effective.


Simply saying, "Ask the necessary questions before answering," will have AI list unclear points.
As a result, you can also organize your requirements, reducing the "hell of re-asking".


9. Assume Iterative Improvement (Rewrite Instructions)

Rather than aiming for perfection in one go, it's faster to assume "draft→revision."
The following format is convenient for revision instructions:

  • What you dislike (Abstract, long, weak, stiff)

  • What you want to do (Concretize, shorten, strong conclusion, add examples)

  • What to keep (Keep this metaphor, maintain the structure)


10. Let It Self-Verify (Check for Mistakes, Omissions, Assumptions)

Particularly emphasized on social media is the problem of "AI casually mixing in plausible lies."
This can be significantly improved with prompts.

  • "List three weaknesses of your answer"

  • "Separate facts from assumptions"

  • "What cases would break the assumptions?"

  • "Clearly state uncertain points as uncertain"


Simply adding this "self-audit" at the end reduces risky assertions.


Social Media Reactions: Praise and Caution Are Discussed Together

When the ZDNET article was shared on social media, reactions largely split into two camps.

 


A. "Ultimately, It Depends on the Prompt" Camp

On Reddit, regarding the quality of AI responses, there were replies suggesting "Isn't it just the way you're asking?" with links to prompt improvement articles as "evidence." In other words, even if there are unsuccessful experiences, there is a sense that there is room for improvement on the user's side.


B. "Blind Acceptance Is Dangerous, Search and Cross-Check Are Essential" Camp

In the same thread, there were repeated warnings that AI might propose something that could break the system or suggest dangerous commands due to hallucinations. The conclusion was that a practical approach is "AI chat + web search + double-check."


C. Shared on LinkedIn as "Work Techniques"

On LinkedIn, the ZDNET article was introduced as a "prompt habit to improve results, assuming AI might fabricate stories," and it was spread in the context of business use (efficiency and quality stabilization).


Conclusion: The 10 Tips Are "Techniques to Reduce Redoing" Rather Than "Speed"

What the ZDNET article refers to as "quick and good results" is not about shortening typing but about reducing the number of re-dos.
By concretizing, providing assumptions, specifying formats, breaking down processes, and finally self-verifying, ChatGPT shifts from being a "convenient toy" to a "reliable partner in practical work."


At the same time, as social media indicates, AI is not omnipotent. Even if it becomes smarter with prompts, human verification is ultimately necessary. Those who don't forget this premise seem to benefit more from AI.


Source URL

Related Articles