I Tried Different AI Prompt Tips for a Week and Here’s What Worked

byAditya Nurrohman -
0

I used to think prompting was just about asking the right question. A week ago, I could write a vague prompt like “Summarize this” and get a mediocre response, then spend another ten minutes rephrasing it, getting frustrated, and finally getting something usable. It felt like a game of chance. I honestly thought the model just had off days. But then I decided to run a structured experiment. I took a week and committed to using only the most recommended techniques from top engineering sources. I banned myself from the generic "improve this" approach and instead treated every interaction like a programming task.

This isn't just a list of tips; this is my diary of what happens when you stop guessing and start engineering.

"A prompt is not a question. It is a call to action. We are in a sense programming with words." - Aditya

The Turning Point: From Chaos to Structure

By Wednesday, the frustration was gone. I stopped treating the AI like a search engine and started treating it like a junior developer who needs explicit instructions. The breakthrough wasn't a single trick, but a system. I realized that the gap between a useless output and a perfect one almost always comes down to how I structured my request.

The most significant shift was adopting the CLEAR Framework [3]. It forces you to move beyond a single sentence. Instead of "Write a blog post," you build a blueprint. You Clarify the objective, Limit the scope, Establish context, Apply constraints, and Review the result.

It sounds rigid, but it’s liberating. Suddenly, the AI wasn't guessing what I wanted. It was executing a plan.

Here is the old workflow I used to follow versus the new engineering approach that actually worked:

FeatureMy Old "Magic 8-Ball" MethodThe New Engineering Approach [1][3]
FoundationVague question or demand.Structured prompt with Instructions, Context, and Output Format.
TechniqueZero-shot (just asking).Role-based + Few-shot examples.
VerificationHoping for the best.Built-in self-review step ("Check your work").
ResultGeneric, often missed the mark.Specific, formatted, and reliable.
Iteration Time10-15 minutes of re-prompting.< 2 minutes of refinement.

The "Persona" Shift: Why Acting Matters

On Thursday, I tested the Role-based (Persona) Prompting technique [1]. The difference was startling. I needed a summary of a complex technical document.

My old prompt: "Summarize this document." (Result: A bland, paragraph-heavy wall of text).

My new prompt: "You are a Senior Tech Blogger for a general audience. Your goal is to explain complex topics in simple, engaging analogies. Summarize this document into three key takeaways using a casual, storytelling tone."

The output was transformed. It used analogies I hadn't even thought of. It adopted the exact voice I needed. I learned that defining a persona isn't role-play; it's focus restriction [9]. You are telling the model exactly which corner of its massive knowledge base to pull from.

The "Show, Don't Tell" Moment

Friday was about Few-Shot Prompting [1]. I needed the AI to categorize customer feedback, but my zero-shot attempts were inconsistent. So, I stopped telling it what to do and started showing it.

I provided three examples of input feedback and exactly how I wanted it categorized. Then, I gave it a new piece of feedback to process.

The result was a night-and-day difference in consistency. The model wasn't just guessing patterns anymore; it was following a template I had explicitly provided. This is the secret sauce for consistency—giving the model a "prototypical" example of the task [1].

The "Tree of Thoughts" Deep Dive

Saturday was the day I pushed the limits. I had a complex strategic problem to solve. Instead of asking for a single answer, I experimented with Tree of Thoughts (ToT) prompting [1].

Instead of one linear path, I asked the AI to act as a committee of three experts: a Financial Analyst, a Risk Manager, and a Growth Strategist. I asked them to debate the problem step-by-step, critiquing each other's ideas before arriving at a final recommendation.

It felt like watching a brainstorming session unfold in real-time. The output was significantly more robust than any single attempt. Research shows that while standard prompting might yield a 7.3% success rate on complex reasoning tasks, ToT can skyrocket that to 74% [1]. It forced the model to explore multiple branches of thought and self-correct, which is exactly what you need for high-stakes decisions.

The "Continuous Improvement" Loop

Sunday, I stopped treating prompts as one-off commands. I adopted a Continuous Improvement Loop [4]. I used a two-agent system: one agent generates the draft, and a second agent acts as a strict critic, evaluating the draft against specific criteria and sending it back for revision.

This is where the magic happens. You stop regenerating blindly. You build a feedback loop. It’s slower and costs more in tokens, but the fidelity is unmatched. As one source noted, if you want a "refined one-time cost" result, this is the way to go [4]. I used it to draft a critical project proposal, and the final version was bulletproof because it had already survived a simulated internal review.

The Verdict: Clarity is the Real Superpower

After a week of rigorous testing, the conclusion is undeniable. Prompt engineering is the act of mastering your own clarity. [6][9]

You don't need to memorize 20 different techniques. You need a framework that forces you to think clearly before you type.

  1. Start with Structure (CLEAR or TCI): Define the task, context, and output format first.
  2. Use a Persona: Give the AI a specific role to narrow its focus.
  3. Show, Don't Tell: Provide examples (Few-Shot) for consistent formatting.
  4. Ask for the "Why": For complex tasks, use Chain-of-Thought or Tree-of-Thought to force reasoning.
  5. Build a Loop: Don't just accept the first output. Use self-review or a critic agent to refine it.

The frustration I felt a week ago wasn't the AI's fault. It was mine. I was speaking in riddles to a machine that only understands explicit instructions. Now, I speak the language of code, just with words.

"Garbage in, garbage out. The AI can only be as clear as the instructions it gets. It forces this incredible intellectual discipline on you." - Aditya

References

Tags

You may like these posts

Post a Comment

Previous Post Next Post