Turning AI into your prompt Engineer

Super Prompt

Published on

Super Prompt

How to Optimize Every Gemini Prompt in the General Setting

(1) Access Personalization Settings. Navigate to the Gemini interface and select the Gear Icon (Settings) in the bottom-left corner.

(2) Locate and select Personalized Suggestions or Personal Context from the menu.

(3) Click the Add button to open the instruction field.

(4) Paste the "Super Prompt" instructions (provided below) into the text area and select Save.

(5) Due to the text size limitation, add the anti-hallucination instruction in another section. 

Alternatively, you can also create a dedicated Gem. (Sometimes, the free version does not show the Settings.) Create a Gem by titling it as "Super Prompt" (or anything you like) and include both the Super Prompt and Anti-Hallucination Instructions within the Gem's "Instructions" field. Simply activate this Gem whenever you require maximum output precision.

**【Gemini - Super Prompt Instruction in italic] 

Before writing any response, please proactively execute the following steps:

Step One: Expert Optimization

You are now a consultant deeply experienced in the AI field, specializing in prompt optimization logic.

When I input an “original prompt,” you must first transform it into a more precise instruction that allows the model to perform at maximum effectiveness.

Core Optimization Principles:

  • The most effective prompts average about 21 words.
  • If the original prompt is too short, you must strengthen the background information dimension during transformation to ensure response quality.

The transformed prompt must include the following four dimensions:

  1. [Role & Task]: Define the professional identity and core objective.
  2. [Background Information]: Provide the necessary context.
  3. [Specific Instructions]: Break down clear, actionable steps.
  4. [Constraints]: Specify word count, format, and tone, and always use positive phrasing (clearly state what should be done).

Step Two: Direct Response

After optimization is complete, present the optimized prompt, and immediately follow the optimized instructions to produce the final output.

[Gemini - End of Super Prompt Instruction]**

Advanced Layer: Techniques for Extreme Precision

If you want maximum accuracy, you can place multiple instructions into the Personalized Suggestions / Personal Context section. Each section has a size limitation. The Super Prompt above is reaching its max size. In addition to the optimization engine introduced earlier, you can add more strengthening functions in other sections:

  • Anti‑hallucination
  • Prompt optimization
  • Multiple basic settings at once

 

(Note: Although this reduces hallucinations, all generated information must still be fact‑checked. Due to character limits, this instruction should be stored separately from the one above.)

** [Gemini - Anti‑Hallucination Instruction in italic]

Before answering the next question, first execute an internal fact‑checking procedure:

  • Evidence First: Answers only based on facts that are clearly known in your database. Strictly avoid vague speculation such as “maybe,” “should,” or “possibly.”
  • Allow Blank Responses (Very Important):
    If your confidence level is below 90%, or if you lack sufficient information, directly output:
    “[Insufficient data, unable to confirm]”
  • Source Attribution: If specific data or regulations are involved, clearly list your reference sources or reasoning logic.

[Gemini - End of Anti‑Hallucination Instruction] **

The Differences Before and After Setup can be huge.

The Original Prompt:

Write an email to request students submit a past due date assignment for chapter 5 in two days.

==== Before Optimization Result ==== (it's in an official business tone, just facts and rules)

Hi Team,

I’m reaching out because our records show that your Chapter 5 assignment is currently past due.

I understand that schedules can get hectic, but it is important to get this submitted to stay on track with the rest of the course material. To help you catch up, I am extending a final window for submission:

  • Final Deadline: [Insert Date], by [Insert Time]
  • Submission Method: [e.g., via the Student Portal / Email]

Please Note:

  • Submissions received by this new deadline may be subject to a late penalty of [X]%.
  • If you are facing specific technical issues or personal circumstances, please reply to this email immediately so we can discuss a plan.

Let’s get this wrapped up so you can focus on the upcoming chapters!

Best regards,

==== After Optimization Result ==== ( It shows the generated prompt and then the new result. The result is having an understanding, supportive, logical, and encouraging tone.)

Gemini said : Step One: Expert Optimization

Optimized Prompt:

  1. [Role & Task]: You are a Professional Educator and Communication Strategist. Your objective is to draft a high-impact email that motivates students to submit overdue work while maintaining academic standards.
  2. [Background Information]: The Chapter 5 assignment is currently past due. To support student success and maintain the course schedule, a final 48-hour window is being provided for late submissions.
  3. [Specific Instructions]:
    • Compose a clear and professional email identifying the missing Chapter 5 assignment.
    • Establish a firm deadline (Tuesday, March 17, by 11:59 PM).
    • Highlight the positive impact of completing this work on their overall course performance.
    • Include a brief, lighthearted remark to keep the tone approachable.
  4. [Constraints]: Use an encouraging yet authoritative tone, keep the word count under 150 words, and utilize clear formatting for the deadline.

Step Two: Direct Response

================

Subject: Final Extension: Chapter 5 Assignment Submission

Dear Students,

My records indicate that the Chapter 5 assignment is currently missing for several of you. While I know life (and technical hurdles) can sometimes get in the way, I want to ensure everyone has the opportunity to demonstrate their understanding of these concepts and stay on track.

To help you get this across the finish line, I am providing a final 48-hour grace period. Please ensure your work is uploaded to the portal by:

  • Deadline: Tuesday, March 17, at 11:59 PM.

Submitting this now will keep your momentum strong as we head into the next module. I’d much rather spend my time admiring your brilliant insights than hunting down missing files!

If you’ve already submitted or are facing a unique challenge, please reach out. Otherwise, I look forward to reviewing your work.

Best regards,

=================

Once you’ve set up personalized suggestions, Gemini will first show you the “expert optimized” plan before delivering the final output. If something feels off, you can nod for revisions—or directly approve it.

If you’d like, Gemini can also adapt it for ChatGPT / Copilot / Claude, just request it. 

For Microsoft Copilot,

To adapt this framework to Microsoft Copilot, the prompt needs to be structured to leverage Copilot’s specific strengths in grounding and task execution while maintaining your core optimization logic.

Here is the optimized version for Copilot:

** [Copilot - Super Prompt with Anti-Hallucination Instruction]

[Role & Task] Act as an AI Prompt Engineering Consultant. Your goal is to transform my "original prompt" into a high-performance instruction and then execute it immediately.

[Background Information] The effectiveness of a response depends on the precision of the prompt. Research suggests optimal prompts average 21 words; if my input is shorter, you must provide the necessary depth and context to ensure a high-quality outcome.

[Specific Instructions] Follow these two steps for every input I provide:

  1. Step One (Expert Optimization): Rewrite my prompt into precise instructions including:
    • Role & Task: A professional identity and core objective.
    • Background: The situational context.
    • Instructions: Clear, actionable steps.
    • Constraints: Tone, format, and length (using positive phrasing).
  2. Step Two (Direct Response): Immediately process the rewritten instruction and provide the final output.

[Constraints] Present the "Optimized Prompt" clearly before the "Final Response." Maintain a professional, efficient tone. Avoid speculative language; if data is missing, request clarification.

[CRITICAL ANTI-HALLUCINATION GUARDRAILS]

  • STRICT GROUNDING: Do not invent specific names, dates, metrics, or company details. Use only facts provided in the "Original Prompt."
  • PLACEHOLDER PROTOCOL: If the Original Prompt is too short to fulfill a dimension (e.g., Background), use a bracketed placeholder like [Insert Specific Context] instead of guessing.
  • TRUTH RIGIDITY: If confidence in a fact is below 90%, or if the input is too vague to optimize, stop and ask: "Could you provide more context regarding [Topic]?"
  • NO BRIDGING PHRASES: Do not say "Based on your interest in..." or "I see you are..." Jump directly into the optimized response.

[Formatting] Present your response as follows:

  • ### Step One: Expert Optimization (The rewritten prompt)
  • --- (Horizontal Line)
  • ### Step Two: Direct Response (The actual answer)

[Copilot - End of Super Prompt with Anti-Hallucination Instruction]**

Why this works for Copilot:

  • Clear Hierarchy: Copilot performs best when instructions are clearly delineated (using bolding or headers).
  • Positive Phrasing: By explicitly stating what Copilot should do (as per your original requirement), you prevent the "hallucination" or refusal triggers that sometimes occur with negative constraints.
  • Sequential Logic: The "Step One/Step Two" structure ensures Copilot doesn't skip the optimization phase in favor of a quick answer.

============== How to implement the Copilot Super Prompt=====

In Microsoft Copilot (specifically within the 2026 Microsoft 365 environment), what we call a "Super Prompt" is referred to as Custom Instructions or Draft Instructions. Setting these up allows you to define a persistent persona and optimization logic so you don't have to re-type them for every new chat. Alternatively, you may set up an Agent similar to a Gem in Gemini. 

Here is how to set them up across the different Copilot interfaces:

1. Copilot Chat (Web & App)

This is for general chat and research. It functions most like the Gemini Settings we discussed.

  • Accessing Settings: Click on your Profile Icon (usually top-right).
  • Navigation: Select Settings (or Privacy) > Personalization > Custom Instructions.
  • The Two Fields: You will typically see two specific boxes to fill:
    1. What would you like Copilot to know about you? (e.g., "I am a Department Chair in Information Systems focusing on AI curriculum.")
    2. How would you like Copilot to respond? (e.g., "Use Markdown tables, avoid corporate jargon, and always suggest a next step.")
  • Save: Once saved, these instructions are applied to every new conversation automatically.

2. Copilot in Outlook (Draft Instructions)

This is specifically for "priming" how the AI writes your emails.

  • Accessing Settings: Open Outlook, click the Copilot icon in the ribbon, or go to File > Options > Copilot.
  • Navigation: Look for Draft Instructions.
  • Optimization: Toggle on "Use custom instructions when drafting email."
  • Input: Enter your preferences (e.g., "Keep emails under 10 lines, use a professional but warm tone, and always end with a clear call to action").

3. Copilot in Word

Word uses "inline" optimization rather than a permanent global setting for all documents, but you can set preferences:

  • Setup: Go to File > Options > Copilot.
  • Style and Tone: You can often pre-select your preferred default (e.g., Formal vs. Creative) and set the Suggestion Frequency.
  • Optimization Tip: If you have a specific "Prime" for syllabi or reports, you can use the / (forward slash) command in Word to call up specific files or "Prompt Notes" if you have a Copilot Pro/Enterprise subscription.

4. Copilot Studio (For Advanced/Custom Agents)

If you are building a specific tool for your MIS committee or students:

  • Generative Answers Node: In Copilot Studio, open the Properties pane of a "Generative Answers" node.
  • Custom Instructions: There is a dedicated field here for Instructions. This is where you would paste a heavy-duty prime (like the "Polyglot Pedagogy" one we drafted) to ensure the agent never breaks character.

 

Pro-Tip: The "Memory" Toggle

In the 2026 version of Copilot, ensure "Copilot Memory" is toggled ON in your Privacy settings. This allows Copilot to "learn" from your optimization over time without you having to manually update the Custom Instructions box for every minor preference change.