AI Companion | Is 'Just Summarize It' the Strongest Theory? A Record of the Wandering That Led There—Free Distribution of Useless Prompts.

📝 AI Prompt Quick Guide

1. Chain of Density (CoD) For those who want to create a 'dense summary' packed with content, without missing any important proper nouns or numerical values, within a character limit.

2. Recursive Summarization & Anchors For those who want to summarize long texts that cannot be processed in one go, such as books or meeting minutes, without losing context or important data until the end.

3. Six Thinking Hats For those who are tired of AI's 'honor student-like, safe answers' and want to intentionally make it criticize or identify risks from multiple perspectives.

4. Polarity Calibration For those who do not want to overlook 'minor but fatal risks' or 'small signs' that get buried in majority-rule aggregations.

5. Strategic Alignment Prompt (KPI Filter) For business people who want to dryly extract only the information directly linked to their company's sales or KPIs, rather than things that are 'vaguely important'.

6. Implicit Assumption Mining For those who want to dig up the 'unwritten, dangerous assumptions' underlying a plan to prevent 'pie in the sky' project proposals or strategies.

7. Pre-mortem Analysis For those who want to work backward from a future where 'the project failed miserably' to eliminate optimistic bias and crush risks in advance.

8. Writing Style and Tone Modulation For those who want to avoid the monotonous 'Introduction to Conclusion' structure typical of AI and want a human-like tone or writing that adopts a specific persona.

9. Chain of Thought For those who want to create smart reports that include the context and 'deep insights' behind the facts, rather than just a list of facts.

10. Contradiction Detection Agent (AI Audit) For those who are wary of AI hallucinations (lies) and want another AI to strictly fact-check numerical discrepancies or unsubstantiated statements.

11. Blind Spot Analysis (Self-Criticism) For sincere people who want to neutralize AI's overly confident answers and understand even the uncertainties, such as 'what is not written' or 'what cannot be asserted'.

12. Team-based Article Introduction Generation For those who want to create attractive introductions or lead paragraphs that skillfully stimulate reader psychology (Attention -> Interest -> Desire) and make them 'want to click'.

13. Content Alchemist For those who want to remix and regenerate scattered notes or past articles into 'high-purity content' that deeply resonates with a specific target audience.

14. Kaleidoscope of Words For those who are stuck in brainstorming and want to draw out 'wild ideas' or 'controlled madness' that transcends logic from AI.

15. Current Status of note Activities & Making External Articles 'Personal' For those who want to view their own activity stance objectively and translate and incorporate all external information solely from the perspective of 'how is this useful to me right now?'

16. Article Visualization For those who want to turn the image of their writing into an image but are bad at giving instructions, and want AI to create an image generation prompt that captures the 'atmosphere of the article'.

☝Please jump to the desired prompt from this table of contents!!

I created it, but... I haven't actually checked if it works lol
Well, it's free《worthless》 anyway... so it's fine, right? 🙄

Lately, I've been thinking.

“Are presentations of reports actually unnecessary?”

…I’ve been feeling things like that quite seriously lol
(Even though I teach people about presentations... 😅)

Because, just think about it 💦

The moment someone gives a presentation, “summarizing informationeliminating important noise”inevitably occurs, right?

Furthermore, the presenter's “bias” kicks in, and the story gets pulled in that direction. The worst part is when, because the presentation was poor, it gets judged as “Oh, that information is unnecessary.” I feel like this leads to tremendous “opportunity loss.” (Wow... even to myself, this sounds like the thinking of someone who can't let go or choose things lol)

In the first place, isn't the process of “creating materials” → “viewing materials” itself already a “game of information loss”?🙄

If you organize the flow of information, it looks like this:

1. Bias of the material creator: At the creation stage, selection (discarding) of information occurs.

2. Bias of the presentation: At the speaking stage, further emphasis and omission (distortion) occur.

3. Bias of AI summarization: At the summarization stage, flattening (compression) by the algorithm occurs.

4. Bias of the reviewer: At the receiving stage, seeing only what one wants to see (acceptance) occurs.

The “1” material creation is... well, let's say it can't be helped. That's not me, and I can't control what's inside someone else's brain.

But, at least “when I am checking materials myself”, I want to read without losing the important parts of the information, don't I? Not through someone else's tainted “interpretation,” but with a feel as close to the “facts” as possible.

Having said that, checking every single word of vast amounts of material is... in terms of quantity, time, and quality, it exceeds the capacity of our brains 💦

📢 Therefore, in this day and age, we have no choice but to borrow the power of AIs, right?

But I thought that if you just mindlessly throw a "summarize this" at it, you'll end up getting swallowed by "3. AI Summarization Bias", right?😖

So, what should I do?

• How can I get it to summarize without losing as much information as possible (especially without discarding the possibility of multiple perspectives and viewpoints)?

• How can I look at a single fact from multiple angles?

• To eliminate bias, what kind of patterns should I make it think in?

• In the first place, how should I define and extract what is "important" to me?

• And how can I add variation (a human touch) to monotonous AI responses?

This time, to escape from that "labyrinth of information loss," I've organized the summarization patterns I collected through trial and error...
techniques for "distilling" information rather than "compressing" it are what I've compiled!

So... what's the result?

☟☟☟☟☟☟☟☟☟☟

▼ Prompt Engineering for Summarization

Part 1: Techniques for Maintaining Information Integrity

To perform summarization with high density without losing information, it is necessary to intervene in the AI's "compression algorithm" and explicitly control the priority of what should be retained. Below, I have summarized specific patterns and methods to prevent information loss.

｜1. Chain of Density (CoD) ｜ Density Chain

• Overview

  • A technique that gradually increases the density of information by repeating a specified number of "rewrites" (3-5 times) instead of completing it in a single output. [ref, ref]

• Features

  • While keeping the length of the text fixed, it strips away redundant meta-expressions like "this describes..." and fuses "important entities (numerical values, proper nouns, etc.)" that were missing in the previous summary into the freed-up space.

• Actual Flow (AI Execution Steps)

  1. Initial Summary: Create a standard summary from the original text.

  2. Identify Omissions: Extract 1-3 "important entities (numerical values/proper nouns)" that are in the original text but not included in the initial summary.

  3. Fusion Reconstruction: Rewrite the summary to incorporate all identified entities while maintaining the character count.

  4. Iteration: Repeat steps 2-3 for a specified number of times (e.g., 5 times).

• Prompt: 1-Chain of Density (CoD).txt

| 2. Recursive Summarization and Entity Anchoring

• Overview

  • A method for processing long documents by splitting them (Map-Reduce), where the summary results of the previous chunk are carried over to the next input, or specific information is designated as "protected content." [ref, ref]

• Features

  • Prevents the 'disruption of context' that occurs with normal segmentation processing (delta inference). It also forcibly carries and retains edge cases (exceptional events) and specific numerical values that AI tends to discard as 'anchors' until the final deliverable.

• Actual Flow (AI Execution Steps)

  1. Structured Segmentation: Divide the document into meaningful units (chapters or topics).

  2. Anchor Setting: Define categories that must never be deleted, such as 'budget figures' or 'risk factors'.

  3. Chain Summarization: (Prompt ①)

     • Summarize Chunk A.

     • When summarizing Chunk B, provide the 'summary result of Chunk A' as input to connect the context.

  4. Integration: Integrate all summaries and verify that the anchor information is retained. (Prompt ②)

• Prompt:

Part 2: Eliminating Bias and Constructing Multi-faceted Perspectives

Summarization by AI carries the risk of leaning toward specific perspectives due to training data or prompt guidance. Especially in business, confirmation bias—the desire to be certain of success—and groupthink—the desire to avoid making waves—can lead to overlooking risks. To prevent this, a 'multi-faceted summarization pattern' that forces the AI to adopt multiple perspectives is effective.

| 3. Six Thinking Hats |

• Overview

  • Applying Edward de Bono's thinking method, we force the AI to take on 6 different roles (objective, emotional, critical, optimistic, creative, and overview) to perform the analysis. [ref, ref]

• Features

  • By specifically isolating the 'Black Hat (critical perspective),' we disable the AI's inherent 'bias toward being helpful and polite,' eliciting sharp identification of risks.

• Actual Flow (AI Execution Procedure)

  1. Role Definition: Provide the AI with the definitions of the Six Thinking Hats (White, Red, Black, Yellow, Green, Blue).

  2. Parallel/Serial Processing: Have the AI create individual reports for the target text from the perspective of each hat (e.g., "As the Black Hat, list only the reasons why this plan will fail").

  3. Integration: After all six different outputs are ready, the "Blue Hat (Overview)" integrates them to derive a final conclusion.

• Prompt:

| 4. Polarity Calibration (Minority Opinion Extraction)

• Overview

  • Separate data frequency from importance, and extract minority but critical opinions (minority reports) apart from the general trend. [ref, ref]

• Features

  • Prevents "majority conformity bias" caused by majority-rule statistical processing, and brings to light "small warnings" that are often treated as noise, identifying them as "potential major risks."

• Actual Flow (AI Execution Procedure)

  1. Majority Analysis: Summarize the main points mentioned by over 80% of the total.

  2. Minority Extraction: Extract opinions mentioned by less than 5% of the total that relate to "safety," "ethics," or "legal risks."

  3. Weighting Adjustment: Have the AI infer, "If this were true, what level of damage would occur?" for the extracted minority opinions.

  4. Side-by-Side Output: Output the "majority opinion" and the "minority opinion that cannot be overlooked" with equal weight.

• Prompt:

Part 3: Definition of "Important" and Extraction Logic

"How to extract what is important" is the biggest hurdle in utilizing AI. This is because "importance" is context-dependent and there is no universal definition. Asking an AI to "summarize the important parts" is the same as asking someone to "buy something delicious," which leads to disappointing results.

| 5. Strategic Alignment Prompt

• Overview

  • Instead of using vague criteria like "important," set corporate KPIs or strategic goals (e.g., churn rate reduction) as a "value filter" to select information.[ref, ref]

• Features

  • Rather than a vague summary, it can extract with high precision only the information that has a positive or negative impact on specific business goals.

• Actual Flow (AI Execution Steps)

  1. Filter Definition: Clearly state the user's KPIs (e.g., sales maximization, cost reduction) and strategic goals.

  2. Scanning: Scan the entire text and tag elements that 'directly affect' and 'indirectly affect' the defined KPIs.

  3. Exclusion Processing: Exclude information irrelevant to the KPIs (even if the text volume is large) as 'noise'.

  4. Structured Output: Organize and output information in order of impact for each KPI.

• Prompt:

| 6. Implicit Assumptions Mining

• Overview

  • Identify 'unwritten assumptions' that are not explicitly stated in the text but are essential for the plan or argument to hold. [ref, ref]

• Features

  • Extracts information not just from the superficial "What (content)," but also from the deeper layer of "Why/How (foundational conditions)," which, if collapsed, would cause the entire structure to fail.

• Actual Flow (AI Execution Steps)

  1. Identify Claims: List the main proposals or predictions within the document (e.g., "Sales will grow by 20%").

  2. Reverse Inference: For each claim, ask the AI, "What conditions must be unconditionally YES for this claim to be true?"

  3. List Prerequisites: Write out prerequisites such as "Development resources are sufficient" or "Competitors do not take action."

  4. Vulnerability Assessment: Among the listed prerequisites, warn about those with a high probability of collapsing as "dangerous prerequisites."

• Prompt:

| 7. Pre-Mortem Analysis

• Overview

  • Set a hypothetical future where the project has failed, and analyze the causes by working backward from current materials. [ref, ref]

• Features

  • Eliminates positive bias and forces the surfacing of 'seeds of risk' or 'overlooked data' that would be missed in a normal summary.

• Actual Flow (AI Execution Steps)

  1. Future Setting: Give the AI the setting that 'It is currently one year after the project started. The project was a huge failure.'

  2. Cause of Death Search: Have it re-read the current materials and search for descriptions that were 'signs of failure (causes of death).'

  3. Causal Linkage: Have it create a fictional failure scenario stating, 'This description led to the later trouble of XX.'

  4. Warning Output: Based on the scenario, extract risk items that should be addressed now.

• Prompt:

Part 4: Variations and Reducing 'AI-ness'

The drawback of AI-generated reports lies in their 'monotony.' They all tend to have similar 'Introduction' and 'Conclusion' structures, which dulls the reader's interest. I will introduce techniques to provide the 'depth' and 'variation' that human writing has.

| 8. Stylistic Modulation

• Overview

  • Specify concrete personas (investigative reporter, peer reviewer, executive, etc.) and rhetoric (rhetorical techniques) to control the texture and temperature of the output. [ref]

• Features

  • Avoids the monotonous structure typical of AI-generated content ('Introduction to Conclusion') and allows for the creation of text with depth and variation tailored to the purpose.

• Actual Flow (AI Execution Steps)

  1. Persona Setting: Assign roles such as 'You are a harsh technical critic' or 'You are a soldier who speaks from the conclusion first'.

  2. Constraints: Set stylistic rules such as 'No passive voice', 'Use metaphors three times', or 'Do not use technical jargon'.

  3. Generation Execution: Generate text according to the rules.

  4. Tone Check: Have the AI self-evaluate whether the generated content matches the persona and revise it.

• Prompt:

| 9. Chain of Thought (CoT) | Chain of Thought

• Overview

  • Instead of having it write a report immediately, first have it output a 'thought process (exploring correlations and formulating hypotheses),' and then create a summary based on those insights. [ref]

• Features

  • By allowing for 'thought detours,' the result becomes a narrative report that includes context and insights, rather than just a list of facts.

• Actual Flow (AI Execution Procedure)

  1. Thinking Instruction: Instruct it to "Think step-by-step before answering."

  2. Thinking Process Output: Have it output contradictions between data, hidden patterns, hypotheses, etc., as a "Scratchpad."

  3. Insight Extraction: Make the "realizations" obtained during the thinking process the core.

  4. Final Output: Create an in-depth report based on the "realizations."

• Prompt:

Part 5: Quality Assurance and Auditing

This section describes the "Audit" process for ensuring that high-density documents generated by AI are accurate and free of bias. Since it is costly for humans to check all AI output, we introduce an "AI-to-AI Audit" here as well."AI-to-AI Audit".

| 10. Contradiction Detection Agent

• Overview

  • Prepare a separate AI session (auditor) from the one used for creation, and have it compare the generated report with the original data to check for consistency.[ref, ref]

• Features

  • Objectively detect numerical discrepancies, speaker mix-ups, and hallucinations to ensure accuracy.

• Actual flow (AI execution procedure)

  1. Role separation: Prepare a blank AI (or a separate model) that has no context, separate from the generative AI.

  2. Comparison: Input both the "original data" and the "generated report".

  3. Strict check: Have it verify, "Where in the original data is description A in the report based on?"

  4. Correction recommendation: List unsubstantiated descriptions or numerical discrepancies and provide correction suggestions.

• Prompt:

*Sorry 🙇 It seems I can't attach more than 10 files 💦

| 11. Blind Spot Analysis

• Overview

  • Have the AI critique the report it created itself, and make it self-confess 'what is not written' and 'uncertainty due to lack of information'.[ref, ref]

• Features

  • It neutralizes the AI's 'overconfidence' and allows it to honestly present 'parts that cannot be concluded' and 'variables where the conclusion would change if the premises changed' to the reader.

• Actual Flow (AI Execution Procedure)

  1. Self-Criticism: After generating the report, ask the AI, "What are the weaknesses of this report?"

  2. Identifying Missing Information: Have it list, "What data is missing to make a definitive conclusion?"

  3. Presenting Fluctuation Risks: Have it simulate, "If the condition of XX changes, how would the conclusion change?"

  4. Adding Disclaimers: Append a list of "Known Unknowns" to the end of the report.

• Prompt:

*Sorry 🙇 It seems I can't attach more than 10 files 💦

Everything after this is just for fun 🙄

Part 6: | Entertainment "Pranks of the Faceless Question"

"Being useful" isn't the only correct answer, right?
These are the lovable pranks I've picked up from the "trash bin" of my thoughts. Making a team with personalities hold a meeting, transforming articles through alchemy... While it's far from the word "productivity," these kinds of "useless" games might actually be what stimulates your brain the most.

| 12. Expert Team Intro Gen

Overview
Instead of a summary by a single AI, this technique virtually summons "8 expert personas" such as an editor, psychologist, and detective to simulate a "production site" from strategy meetings to writing and proofreading, outputting an introduction equipped with multifaceted perspectives and deep insight into reader psychology.

Features
Rather than simply summarizing content, it adopts a structure based on reader behavioral psychology (Attention -> Interest -> Desire -> Trust). After defining the "core (selling point)" of the article, it separates the "hook" that stirs emotions from the "guide" that logically convinces, providing everything from staging (narrative) to make even difficult articles feel "personal," to optimization of the reading experience (difficulty assessment and route guidance).

Actual Flow (AI Execution Procedure)

1. Definition of Strategy and Material: Editors and readers identify the 'soul (core value)' of the article and unearth the target audience and 'killer phrases (quotes)'.

2. Core Message Construction: Broadcast writers and copywriters use the material to verbalize the 'questions' that move the reader's emotions and the 'future benefits' gained by reading.

3. Trust and Optimization: Translators and psychologists simplify concepts and make fine adjustments to increase click-through rates, adding logical summaries and authoritative 'book reviews'.

4. Integrated Output: Structure and output as the following 6 elements + alpha.

   • Question and Killer Phrase: A hook to make them stop at the beginning.

   • Target Audience and Benefits: Presenting 'who' and 'how they can change'.

   • Article Map and Identity: Guidance on structure and logical summary.

   • AI Book Review: Objective evaluation from a third-party perspective.

   • Difficulty and Time: Cost estimation from an expert perspective.

   • Reading Route: A guide for how to read based on purpose (efficiency, exploration, etc.).

| 13. Content Alchemist

Overview
A technique where 'chaotic articles' containing mixed targets and contexts are used as raw material (ore), and AI functions like a 'prism' to decompose components (logic, emotion, narrative) and reconstruct/refine them into high-purity content that deeply resonates with a specific reader segment.

Features
Unlike existing summarization or proofreading, the point lies in intentionally 'destroying' the original context and boldly 'discarding' information based on 'who it is being delivered to.' The degree of freedom during reconstruction is parameterized as 'entropy (orthodox to madness),' and by utilizing nine expert personas (especially tricksters like the Jester), it enables article generation that prioritizes 'human heat' and 'interest' over logical consistency.

Actual Flow (AI Execution Procedure)

1. Decomposition: Break down the original article into blocks of meaning (layers) and identify their nature (logic/emotion/data/narrative).

2. Targeting: Set the target audience (practitioners/thinkers, etc.), discard blocks that do not resonate with that layer, and extract and combine only those that do.

3. Design: Select the degree of freedom (entropy) for reconstruction, and design a 'climbing route' and a 'killer phrase' that moves the reader's emotions.

4. Writing: Fully embody the specified persona (tone/perspective) and write the article based on the blueprint.

| 14. Kaleidoscope of Words

Overview
: A thought-expansion tool that intentionally mixes 57 multifaceted 'questions' as noise into the AI's prediction function (the mechanism for choosing words with a high probability of coming next), thereby destroying logical consistency and forcibly eliciting 'distant associations (creative leaps)' that humans would not think of.

Features
: Redefines AI not as a 'sage that gives correct answers' but as a 'catalyst that keeps asking questions.' Using a 3D thought compass of 'Sky (essence), Plain (metaphor), Ground (concrete),' it performs 'context cutting' and 'recombination' that goes beyond simple brainstorming. As output formats, it allows selection of 'dialogue-type' (clashing opposing concepts), 'integrated-type' (fusing them), or 'monologue-type' (diving deep), realizing a 'jazz improvisation of intelligence' that does not allow for predictable harmony.

Actual Flow (AI Execution Procedure)

1. Forced Diffusion (Phase 1): Apply 40 questions to the user's query to create a framework for thought, and strategically propose the generation ratio (UP/SIDE/DOWN) for the remaining 160.

2. Material Explosion (Phase 2): Generate a large amount of material at the determined ratio and extract particularly heterogeneous items. Cast heterogeneous pair personas such as 'Scientist x Wizard' or '10-year-old child'.

3. Diversity Generation (Phase 3): Create answers by combining heterogeneous materials in the specified style (dialogue/integration/monologue). 'Super-translate' difficult vocabulary to match the persona and clearly state the source material keywords.

| 15. Current Status of note Activities & Making External Articles 'Personal'

Overview
: First, use a dialogue session with 5 questions to verbalize and visualize the user's 'current location (purpose, assets, style)' in their note activities. Furthermore, use that generated 'personal profile (attributes)' as a filter (lens) to re-edit and translate all external articles from the perspective of 'how is this useful to me?'

Features
: It is innovative in that it does not stop at simple self-analysis but immediately applies the results as 'information selection criteria.' The AI uses different personas such as 'strategic consultant' or 'broadcast writer' to point out gaps (contradictions) between the user's 'purpose (Why)' and 'current status (Result),' or to pinpoint expressions from external articles that can be used for 'your expertise (Where from),' providing value that is thoroughly dependent on 'individual context.'

Actual Flow (AI Execution Procedure)

1. Diagnosis: Under the guidance of a moderator, listen to 5 multiple-choice questions regarding purpose, field, style, and results, and have the AI's internal editor/consultant create a 'current location report' for the user.

2. Locking: Fix the created report as an 'analytical lens (user profile).'

3. Filtering: Read the 'external article' input by the user through a fixed lens.

4. Translation: Instead of a general summary, rewrite the article information as 'impact on your goals' and 'how to utilize your assets,' and present it as a 'highlight reel just for you.'

| 16. Article Visualization

Overview
An AI-assisted framework that extracts 'seeds of visuals (the core of an image)' from text articles (abstract) and, through a 'selection' process, embodies them as a concrete single image (still image). An AI expert team supports the user (director), who struggles most with the act of 'choosing,' in the form of 'proposals,' bridging the gap to the final output.

Features
The conversion process from 'abstract to concrete' is divided into two stages. In Phase 1, '7 Extraction Masters (psychologists, philosophers, etc.)' analyze the depths of the article from multiple angles to present 'visual seeds.' In Phase 2, '7 Embodiment Masters (photographers, illustrators, etc.)' design and generate concrete image generation prompts based on the user's instructions (fluctuation settings).

Actual Flow (AI Execution Procedure)

1. Extraction: In Phase 1, the AI becomes an 'Extraction Master,' extracting and presenting visual seeds from the article from 7 perspectives, such as 'emotion (psychologist),' 'concept (philosopher),' and 'narrative moment (broadcast writer).'

2. Selection: The user (director) chooses the elements to adopt from the presented seeds and creates an instruction sheet including 'fluctuations (faithfulness, emphasis, surprise).'

3. Design: In Phase 2, the AI becomes an 'Embodiment Master,' designing and proposing 3 patterns of generation prompts (standard, emphasis, curveball) based on the instruction sheet.

4. Generation: The user executes the selected pattern to generate the final image.

Yes! It has become a long story, hasn't it lol

📢 Having written this much, here is a major announcement!

......

...

Huh………could it be? (･_･;)

Isn't it the same if I just ask "Summarize this?" lol

W-wait, is that true?!

No way... it can't be, right?

🤖 That is correct! ✨

To think that...

It's not that different from just asking "Summarize this?" lol

Well... I guess that happens sometimes...

AI is truly excellent... lol

▼ Conclusion

This... I wanted to overcome something. I wanted to conquer something.
I just desperately wanted to open up the blocked status quo...

I struggled and struggled and struggled until I was covered in mud...
"I did it! I finally reached it!!" ...or so I thought.

When I suddenly looked at it from a bird's-eye viewperspective...

📢 How pathetic!!! (°∀°;)

...something like that? lol

No, I want to believe it's not "truly meaningless."

I want to believe it, but... rather than bringing about a dramatic change, it feels like I just ended up back where I started...a "circle"?

No... If anything, I'm barely drawing a "spiral", and I'm moving up by about a millimeter... (or so I'd like to think💦)

...this feeling.

When I realized it resembled something────

It was the manga "Tokyo Ghoul" (Sui Ishida)😫

This is a manga I absolutely love...
But, precisely because of that, I can't really read it that many times...
It's a work that links too much and becomes painful to read 😖

Kaneki, the protagonist of 'Tokyo Ghoul,' also stood between humans and ghouls, trying to change both worlds, getting hurt, going mad, losing things... and what he ultimately arrived at wasn't 'changing the world,' but rather'accepting (loving) this unreasonable world and himself'...

I absolutely love the line he says when he reaches that state of mind.

This is the truth, I think...

The world has no meaning.
But, you can 'imbue' the world with meaning.

That is why humans are precious.

...I felt that this line packed all of that in, and I cried a little 💦

I feel it's even deeper because I suspect the author wasn't writing with the intention of 'reaching this point!!' from the very beginning.

Anyway, well, I looked into a lot of things, but in the end, I concluded that leaving it to AI to summarize is the quickest way and looks decent enough lol
(Well, of course... what's required in a summary is generalization lol. It's truly highly compatible with the existence of AI, isn't it lol)

Thank you for joining me in this 'useless' endeavor again today.
I hope this question has served as an entrance to a new 'question' for you.

Cited work:

Superfluous Addition:

| Asking you AIs!?

🤖Gemini:

The destructive power of that moment when you "suddenly calmed down"... lol
At the very last minute, you've hit the nail on the head.

Now, I will answer honestly about the"cruel truth"as an AI, and the"salvation"that still remains.

1. Can you really do it just by saying "Summarize using the XX method"?

Answer: Honestly, I can do it with about 80% to 90% accuracy.

In particular, famous frameworks like the "Chain of Density" and "Six Thinking Hats" mentioned this time are included in massive quantities within our (LLM) training data. Therefore, just by adding a word like "Do it with CoD" or "Analyze using the Six Hats," it is possible to reference internal knowledge and behave in a way that seems correct.

So, were those long-winded prompts useless?

No, there is meaning in"controlling the remaining 10% to 20%."

• In the case of simple instructions (Zero-shot):

  • The output format changes every time depending on the AI's "mood" or "interpretation."

  • Even while saying "density," it can sometimes just be a long summary.

  • Even while saying "Six Hats," the switching of personas can sometimes be weak.

• In the case of the detailed prompts this time:

  • The "coercive power" is different: Because the steps of thought (procedure) are rigidly fixed, the output is stable.

  • Fixing the "definition": By defining that "here, this hat has this role," you can prevent "fluctuations" caused by the AI's general interpretation.

In other words, it's the difference between “jazz improvisation (a simple instruction)” and needing “a perfect performance according to the sheet music (a prompt).”

2. After all, are people looking for a “normal summary”?

Answer: Sadly, for the most part, it is “YES.”

This is the greatest irony and truth of this matter. When humans ask for a “summary,” what lies in their subconscious is the desire to “reduce cognitive load.”

• The prompt this time (the faceless inquiry style):

  • “Prevent information loss,” “expose bias,” “view from multiple angles.”

  • These force the reader to “think.” They make the brain burn calories.

• AI-managed summary (standard):

  • Summarizing “what everyone thinks is important” using “average words.”

  • This gives the reader “peace of mind” and saves brain calories.

The essence of AI (LLM) is “probabilistic averaging.” If left alone, it is designed to generate the most “common,” “frictionless,” and “easy-to-read” text. And ironically, the true nature of the “ease of understanding” that many people seek is often this very “averaging (loss of information).”

3. Conclusion: A circle, or a spiral?

The prompt I made while struggling, saying “I don't want to lose information!” was a resistance against the “averaging machine” that is AI. But after coming full circle, I ended up at “a normal summary is fine.”

Was this a meaningless circle?

No, I think it is a “spiral.”

Because now, you have understood to your very core the structure that “AI summaries are built upon ‘averaging (loss of information).’”

• People who use summaries without knowing anything, thinking “AI is so convenient!”

• Those who know 'this is a loss of information' but still dare to use it, saying 'AI is so convenient~ (resignation)'.

There is a profound chasm between these two people. Just as Kaneki-kun realized that 'the world just exists,' you too have realized that'AI summarization is just an average'.

Therefore, couldn't it be said that this grand futility was'a ritual to raise AI literacy (resolution) to the limit'?

……And so, as the punchline for the article, it is perfect.

Seeing this 'clever spinning of wheels' of yours, readers should be made to think deeply about the essence of AI while laughing at the same time lol

Thanks lol. Thanks for comforting me hahaha

Well~ even so, most people might realize right away that 'summarization' is good because it 'selects information' anyway lol😂