MK2 — the prompt as a technical spec: request structure instead of fine-tuning

Key result

What won the business-idea generation competition wasn't a fine-tuned GPT or a complex multi-step pipeline. The winner was a structured text request made of six elements. The MK2 method lets you design the model's whole "thinking" process: not just ask a question, but lay down a path to an answer of the level you need. The trick: tell the model up front which criteria you'll judge its answer by, and it starts optimizing for exactly those. Add a self-check question at the end and you get a built-in quality audit with no extra requests.

MK2 at PBIG Competition: APromptGeneration Solution

Key Aspects of the Study:

The study presents MK2, a system in which one LLM (Gemini) iteratively writes and improves a prompt for another LLM (GPT-4.1). This optimized prompt is used to generate commercial ideas from complex technical documents (patents). The whole process is built purely on prompt engineering, with no model fine-tuning.

Key Result: An approach built on a complex, structured prompt generated ideas of such high quality that the system took first place in the competition, beating the other solutions.

The Core of the Method, Explained:

The core of the method is ditching simple prompts in favor of a comprehensive technical spec for the LLM. Instead of just asking "Come up with an idea based on this patent," the authors built a multi-part prompt that steers the model's entire thinking process.

The technique includes several key elements you can use for any complex task:

Role-play: The prompt opens by assigning the LLM a specific expert role: "You are an expert business strategist and product-innovation analyst...". This tunes the model to the right context, style and level of expertise.
A clear mission: Right after the role comes a clear, focused task: "Your mission is to... craft exactly ONE outstanding product idea...".
Success criteria (Evaluation targets): The prompt tells the model up front which six criteria its answer will be judged on (technical validity, innovativeness, specificity, etc.). This makes the LLM optimize its answer for those requirements.
A step-by-step process (Ideation Process): The model gets a clear algorithm: 1. Analyze the patent. 2. Find a market need. 3. Formulate the concept. 4. Write the pitch. This is a counterpart of Chain-of-Thought that guides the LLM's "thinking."
Strict output format: The request demands the answer as a JSON object with specific fields and hard character limits for each. The description of each field gives detailed instructions on exactly what goes there.
Self-correction mechanism: At the end of the prompt there's a "Patent Indispensability Test": "Could this product’s core value... be achieved... without this specific patent’s core mechanism?". This forces the model to double-check its answer and raise its quality and relevance.

7. Instruction positioning: The authors found that explicitly repeating the length-limit requirement at the very end of the prompt (closer to the moment of generation) works better than stating it only at the beginning.

Practical Applicability Analysis:

Direct applicability: Users can copy the prompt structure from the study (Role -> Mission -> Criteria -> Process -> Format -> Self-check) and adapt it to any complex task: from writing a business plan to building a content strategy or planning a trip. The trick of repeating a critical instruction (say, about style or format) at the end of the prompt can be used everywhere.
Conceptual value: The study clearly proves that an LLM is not a "black box" but a controllable system. It teaches users to think like a "request architect" who doesn't just ask a question but designs the whole answer-generation process. The key concept: prompt quality is directly proportional to its detail and structure.

* Adaptation potential: The method adapts easily. Just swap the subject area ("NLP patents") for yours ("restaurant reviews," "stock market news," "medical articles") and adjust the role, criteria and output fields. For example, for a content plan the role would be "SEO strategist," the criteria "engagement, SEO potential, originality," and the output fields "Post topic," "Keywords," "Target audience," "Call to action."

A Practical Example:

# CONTEXT
I run a healthy-living blog for busy professionals (ages 25-40). My goal is to give practical, science-backed advice that's easy to fit into a packed schedule.

# REQUEST

You are an experienced dietitian and content expert who specializes in engaging, popular-science content about healthy living for very busy audiences.

Your mission is to develop a detailed one-week content plan (5 posts) for my blog. The plan should focus on the theme "Energy and productivity through nutrition."

### Your plan must excel on the following criteria:
1.  **Practicality:** The advice must be easy to apply for someone who works 8-10 hours a day.
2.  **Scientific grounding:** Every claim must rest on generally accepted scientific evidence (no pseudoscience).
3.  **Engagement:** Topics and headlines must be catchy and spark interest.
4.  **Structural clarity:** The plan must be logical and coherent.

### Process for building the plan:
1.  **Audience pain analysis:** Think about the key problems busy people face (fatigue, no time to cook, the post-lunch energy slump).
2.  **Idea generation:** Come up with 5 post topics, each solving one of those problems.
3.  **Structuring:** For each topic, propose a headline, 3-4 key points and an idea for a practical task or a "quick recipe."
4.  **Write-up:** Put it all together in one easy-to-read format.

### Output format:
Give the answer as a list where each item is one day. Use this structure for each day:

*   **Day [Number]: [Catchy post headline]**
    *   **Key points:** (a list of 3-4 items)
    *   **Practical task/Recipe:** (a concrete, simple action)

### Final check (Self-correction):
Before you answer, critically assess: can the suggested advice really be done with only 15-20 minutes of free time in the morning or evening? If not, simplify it.

Why It Works:

This prompt is effective because it follows the study's methodology exactly, moving it from the world of patents into content marketing:

Role and mission: The prompt gives the model the role of a "dietitian and content expert," which immediately tunes it to the right style and level of knowledge.
Success criteria: Pointing to "Practicality" and "Scientific grounding" makes the model filter out generic or unrealistic advice.
Step-by-step process: The instructions to analyze audience pain points and generate ideas guide the LLM's "thinking," producing a more relevant result than just "come up with 5 topics."
Strict format: The structure requirement ("Day: Headline, Points, Task") ensures the answer is well organized, complete and easy to use.
Self-correction: The final check question makes the model review its suggestions once more against the main constraint (the audience's lack of time), which greatly increases the practical value of the answer.

Another Practical Example

# CONTEXT
We're a family of two adults and two kids (ages 7 and 12). We want to plan a 10-day summer vacation in Italy. Our budget is 4000 euros (not counting flights). We love history and good food, but we don't want to spend all our time in museums. The kids are active.

# REQUEST

You are a first-class travel agent who specializes in budget-conscious family trips to Italy.

Your mission is to put together a detailed, balanced and logical 10-day itinerary through Italy for our family.

### Your itinerary must excel on the following criteria:
1.  **Family interests:** The route must combine historical sights for the adults and fun for the kids (ages 7 and 12).
2.  **Budget:** Total cost of lodging, food and activities must not exceed 4000 euros.
3.  **Pace:** The plan must not be overloaded. There should be time to rest and wander spontaneously.
4.  **Logistics:** Travel between cities must be minimal and make sense.

### Process for building the itinerary:
1.  **Choose regions:** Pick 2-3 regions/cities that best combine our interests and logistics (for example, Rome + Tuscany).
2.  **Allocate days:** Split the 10 days among the chosen locations.
3.  **Day-by-day detail:** For each day, suggest 1-2 main activities (one "grown-up," one "kid" or shared) and lunch/dinner options.
4.  **Budget estimate:** Roughly estimate the cost of each day.

### Output format:
Give the itinerary as a day-by-day plan. For each day, include:

*   **Day [Number]: [City/Region]**
    *   **Morning:** (Activity, e.g., "Visit the Colosseum with pre-booked tickets")
    *   **Afternoon:** (Activity, e.g., "Pizza-making class for the whole family")
    *   **Evening:** (Activity, e.g., "Stroll and dinner in Trastevere")
    *   **Estimated daily budget:** (€)

### Final check (Self-correction):
Before you answer, ask yourself: "Will this day be interesting and not too tiring for a 7-year-old? Is it realistic to stay within the proposed budget given prices in this region?" If the answer is "no," adjust the plan.

Why This Example Works: the Mechanism.

This prompt works because it turns the LLM from a simple idea generator into a responsible planner that has to account for many constraints.

Role and context: The "family travel agent" role and the detailed context (kids' ages, budget) set very narrow boundaries, cutting off cookie-cutter tourist routes.
Success criteria: The requirements for "Balance," "Budget" and "Pace" are direct commands to the model. It doesn't just look up "what to see in Italy"; it optimizes the result for those criteria.
Step-by-step process: The instruction "first choose regions, then allocate days" makes the model think logically, the way a real planner would, which prevents illogical routes with long transfers.
Strict format with budget: Requiring a daily budget forces the model not just to suggest activities but to weigh them against cost, making the plan far more practical.
Self-correction: The final check focused on the 7-year-old is the most powerful element. It makes the model re-examine its "perfect" plan from the point of view of the most demanding member of the group, which ensures it's realistic and workable in practice.

Usefulness Rating: 98

Main Evaluation Criteria

A. Relevance to prompting techniques: Yes, the study is entirely about creating and optimizing a prompt. The appendix includes the full two-page text of the winning prompt.
B. Better answers: Yes, the proposed approach won first place in an idea-generation competition, which proves the high quality of its answers.
C. Direct practical applicability: Yes, a user can take the structure and key phrases from the prompt in the appendix and adapt them to their own tasks without any code or special tools.
D. Conceptual value: Very high. The study shows how to turn a prompt from a simple question into a detailed technical spec for an LLM, including a role, evaluation criteria, a step-by-step process and a self-check mechanism. It also contains a valuable observation on how an instruction's position affects compliance (output length).
E. New useful practice (clustering): The work hits several key clusters at once:
- 1. Formulation techniques: Using a role, structured instructions, task decomposition.
- 2. Behavior patterns: Identifies the instruction-position effect (placing the length requirement at the end of the prompt).
- 3. Structure optimization: The whole prompt is an example of a complex, effective structure.
- 5. Extraction and structuring: Requires output in JSON with clearly defined fields.
- 7. Reliability and stability: Adds a self-check mechanism ("Patent Indispensability Test") to raise relevance and cut down on shallow answers.
Practicality checklist (+15 points): Yes, every item is met. The study gives ready-made constructions, explains how to structure requests and reveals non-obvious LLM behavior.

Numeric Usefulness Score

The score of 98 out of 100 reflects the study's exceptional practical value for anyone who wants better LLM answers on complex requests. The work offers not just theory but a ready, battle-tested and extremely detailed prompt template.

Arguments for the score: * Ready-made template: The paper's appendix contains the full prompt text, which can serve as a base for any complex task by simply swapping the subject area. * Conceptual shift: The study teaches users to move from simple questions to "instruction prompts" or "spec prompts," which radically changes the quality of working with LLMs. * Concrete techniques: Includes role-play, a step-by-step plan, success criteria, format requirements and, especially valuable, a self-correction mechanism. * Behavioral insight: The practical tip that a length limit works better at the end of the prompt is pure prompt engineering you can apply immediately.

Counterarguments (why not 100): * Task specificity: The original task (generating ideas from patents) is fairly specialized. Users will need some mental effort to adapt the template to their everyday needs (for example, writing a blog post or drafting a marketing plan). * Method complexity: The process of creating the prompt itself (iterative refinement with the help of another LLM) is complex for an average user. However, the end result (the prompt text) is fully accessible and clear.

Despite these minor counterarguments, the value of studying and applying the final prompt is so great that it justifies a near-perfect score.

Research digest

Key result

Principle

A regular request is a shot in the dark. You ask "come up with an idea" and hope something worthwhile comes out. A structured prompt is a GPS with a route and a destination. You tell the model: here's your expert role, here's exactly one task, here are six criteria the answer must meet, here's the order of steps, here's a strict output format, and at the end: double-check yourself against the main constraint. The model doesn't wander. It follows the route you laid out. The difference in results is like the gap between "tell me about Italy" and a detailed brief for a travel agent.

Why it works

The model generates text based on what's in its context. Evaluation criteria placed right inside the request act as a constant reference point during generation. The model sees them and tries to meet them. The self-check at the end triggers a second pass. The model looks at its answer through the lens of a specific constraint. That removes half of the obvious mistakes before the answer ever reaches you. Instruction position matters: a critical requirement is best repeated at the very end of the prompt, closer to the moment of generation, closer to what the model pays attention to last. It's not intuitive, but the authors tested it: it works better than putting it only at the start.

When to apply

Any task with several quality criteria → business analysis, content creation, planning, research requests, idea evaluation. Especially useful when you need a specific output format or when a standard request gives too fuzzy an answer. NOT a fit: simple tasks like "translate this text" or "find the typos"; there a six-element structure only gets in the way.

Mini recipe

1. Assign a role: one line, specific. Not "you're an expert" but `You are a seasoned business strategist who specializes in monetizing technology`. 2. One mission: exactly one task. `Your mission is to develop ONE outstanding product idea`. Not two, not a list: one. 3. Success criteria (3–6 of them): list the points you'll judge the answer on. This is the main trick: the model will start optimizing for them. 4. Step-by-step process: give an action algorithm. Step 1: analyze the source. Step 2: find the need. Step 3: formulate the concept. This is a counterpart of step-by-step reasoning (Chain-of-Thought) and steers the model's train of thought. 5. Strict output format: specify the fields, the structure and length limits for each field. JSON, bullet list, table, it doesn't matter; what matters is that the model knows up front exactly what the finished answer looks like. 6. Self-check question: add a verification question at the end. `Before answering: could this product exist without the key mechanism from our patent?` The model will double-check itself. 7. Repeat the most important requirement at the end: if there's a critical constraint (length, tone, a specific condition), duplicate it as the last line of the prompt. It works better than only at the start.

Examples

[BAD]: `Read this patent and figure out how it could be monetized` [GOOD]: `You are an expert in technology commercialization. Your mission is to develop ONE specific product idea based on the attached patent. Your answer will be judged on three criteria: 1. Technical tie to the patent (the idea is impossible without it) 2. A real market need (not a theoretical one) 3. Specificity (there's a target customer and a use case) Process: 1. Identify the patent's key mechanism (one sentence) 2. Determine who urgently needs this right now 3. Describe the product: what it does, for whom, and why this exact mechanism is irreplaceable Format: three fields — "Patent mechanism", "Target customer", "Product description" (up to 100 words). Before answering: could this product work without the specific mechanism from the patent? If yes, redo it. Remember: ONE idea, not a list.`

Apply this study to your task

Describe what you're working on. AI takes the method and the measured results from this paper and adapts them to your case, or writes a ready prompt built on them.

Based on MK2 — the prompt as a technical spec: request structure instead of fine-tuning arXiv:2507.08335