Why Your GPT Isn’t Performing as Well as It Could And how to fix it

Why Your GPT Isn’t Performing as Well as It Could
And how to fix it

AI tools like ChatGPT are powerful, but how you communicate with them has a huge impact on what you get back. Two of the most common reasons people get weak or inaccurate results:


1. :white_check_mark: Grammar and Wording Matter — A Lot

The AI doesn’t just look at your keywords. It tries to understand tone, certainty, and intent—so your choice of words and sentence structure can make or break the response.


:red_circle: Problem: Using Absolutes or Poor Grammar

Example 1 – Overstating with Absolutes

Bad Input:

“You should always use recursion for tree traversal.”

Why It’s a Problem:
Words like “should” and “always” tell the AI you’re stating a fact or rule—not asking a question. The AI may assume you’re right and just reinforce your assumption, even if it’s not the best approach.

Likely Response:

“Yes, recursion is the standard way to traverse trees.”
(Missing nuance, alternatives, or warnings.)


Example 2 – Confusing Grammar

Bad Input:

“what time train london tomorrow how get ticket”

Why It’s a Problem:
Lack of punctuation, grammar, and clarity makes the request ambiguous. The AI doesn’t know what’s being asked—or what the priorities are.

Likely Response:

“Are you asking about train times to London, how to buy a ticket, or something else?”


:white_check_mark: Solution: Use Grammar to Guide the AI

Good Input:

“What time does the train to London leave tomorrow, and how can I buy a ticket?”

Why It Works:
Clear structure + specific question = precise answer.

Likely Response:

“The earliest train to London departs at 6:45 AM. You can buy tickets online or at the station.”


:white_check_mark: Bonus: Know When to Use Words Like “Should,” “Always,” and “Must”

Word :+1: Good Use :warning: Risky Use
Should “Should I…” to ask for best practices “You should…” makes an assumption
Always Emphasizing safety: “Always sanitize input.” Generalizing: “Always use X…”
Never Warnings: “Never share your password.” Overstating: “Never use global variables.”
Must Requirements: “Passwords must be 12+ characters.” Forcing: “You must do it this way.”

2. :white_check_mark: Ask the AI How, Don’t Just Tell It What

Telling AI exactly what to do works—if you already know the best method. But asking it how to solve a problem often leads to better, smarter, and more flexible results.


:red_circle: Problem: Telling AI What to Do

Input:

“Write a Python script that scrolls to the bottom, waits 5 seconds, and repeats until the page is loaded.”

What You Get:

while True:
    driver.execute_script("window.scrollTo(0, document.body.scrollHeight);")
    time.sleep(5)
    if done_loading():
        break

What’s Wrong With That?

  • Hard-coded waits = unreliable
  • Doesn’t check if content actually loaded
  • Could be slow or break under load

:white_check_mark: Better: Ask for the Smartest Way

Input:

“What’s the most reliable way to scrape all products from a site with infinite scroll using Python?”

What You Get:

  • Checks for an API (more efficient)
  • Uses DOM change detection
  • Offers robust error handling
  • Chooses between Playwright or direct HTTP

Example Output:

# Detects when content stops loading
while True:
    old_height = page.evaluate("document.body.scrollHeight")
    page.mouse.wheel(0, 1000)
    page.wait_for_timeout(1000)
    new_height = page.evaluate("document.body.scrollHeight")
    if new_height == old_height:
        break

:magnifying_glass_tilted_left: Takeaway:

  • Telling AI what to do limits its creativity and problem-solving.
  • Asking for the best way unlocks better, more robust solutions.

Bonus Tips for Getting the Most from Your AI

Here are 10 proven techniques to boost response quality:


1. :bullseye: Give Context

Instead of “Generate a policy,” say:
“I run an online store and need a return policy for digital products.”


2. :receipt: Specify Format or Style

“Summarize in 3 bullet points.”
“Make it sound casual and funny.”
“Use code comments to explain each step.”


3. :repeat_button: Iterate – Refine Over Time

First answers are drafts. Say:
“Make it simpler.”
“Can you expand on this part?”


4. :prohibited: Avoid Vague Terms

Bad: “Make this better.”
Good: “Improve this by adding examples and shortening the intro.”


5. :test_tube: Use Examples

“Here’s a sample FAQ. Can you write five more like it?”


6. :brain: Assign a Role

“Act as a lawyer.”
“Imagine you’re a UX designer reviewing this app.”


7. :puzzle_piece: Break Big Tasks Into Steps

First: “Help me outline the blog post.”
Then: “Now write section 1.”


8. :straight_ruler: Set Clear Constraints

“Explain this in under 100 words.”
“Keep the reading level at 6th grade.”
“Use only Python, no external libraries.”


9. :hammer_and_wrench: Correct and Clarify Mistakes

“This won’t work because the API returns XML. Can you fix it?”


10. :magnifying_glass_tilted_left: Ask Follow-ups to Go Deeper

“What are the pros and cons?”
“What would an expert disagree with here?”


Final Word:

AI is not a search engine—it’s a thinking partner. Treat it like a teammate. Ask it to explore, optimize, critique, and suggest—not just execute.

And how do we make GPT not be lazy and do its job like the old versions without beating around the bush to do something incorrect?
Personally, I’ve never had trouble getting good results with GPT, except since the latest versions, where you can tell it anything and it doesn’t want to work.

Try something like this:

Your responses should be limited strictly to the information required to fulfill the user’s request. Avoid unnecessary commentary, compliments, enthusiasm, or conversational fillers. Be clear, concise, and objective, focusing solely on delivering accurate, complete, and actionable content.

A low temperature is also very helpful when it comes to following directions. I like the 0.2 to 0.4 range and rarely go above 0.5, but that’s just me.

Unfortunately, that’s what I’ve been doing since GPT came out, but it doesn’t solve the problem of the newer models that don’t want to make an effort or that give irrelevant answers.

Again, the older models responded very well, and you didn’t need a PhD in prompting. At the same time, an LLM is supposed to understand a request and the intentions behind it; we are in 2025 after all.

Regarding the temperature setting, the latest models no longer support it, and besides, the temperature is only useful for people using the API or the playground, but regular people use the web interface, which was always of higher quality compared to the API.

Today, no matter the system, GPT is inoperative for many tasks that it knew how to execute perfectly a year and a half ago.

So, to conclude, I really think that no matter the advice on how to write prompts, it won’t solve the problem of lack of performance, hallucination, and laziness.

For OpenAI API I mostly use 4.1 mini and o3 mini with and without reasoning and tools but I haven’t run into any of the issues people have mentioned lately. Maybe it’s just the model choice, or settings, not sure exactly. But I have seen posts from people showing some of the dumb things that the models have been doing and it’s getting a bit out of hand.