You are free to copy and sell this entire article.
First, let me say the most important thing.
You are welcome to take the contents of this article, copy it entirely, and sell it as your own content.
Whether you repost it on Note, sell it on Brain, incorporate it into course texts, or break it down into SNS posts, it's all OK. No permission or contact is required. Even if you present it as if you came up with it yourself, I won't complain at all. There are about 40 techniques included, so just by extracting one technique per post, you can create 40 pieces of content.
Why go this far? The reason is simple: if I don't, Japanese people won't realize the value of this information.
To be honest, among global AI power users, Japanese prompts are openly called "Stone Age level." When I first heard that, I thought, "What is this guy talking about?" But the moment I actually saw the prompts they use daily, I was speechless. It was a total defeat.
What was the difference? 99% of Japanese people write prompts as "instructions." "Write a blog post," "Summarize this text," "Give me 5 ideas." These are all just commands. On the other hand, global pros pass the "thought process" and "goals" to the AI from the very first line. It's the difference between a command and a blueprint.
And here's a harsher reality. While you are looking for "a nice prompt," global players are researching prompts through academic papers, managing them like code, and automatically optimizing them with algorithms. Handcrafted craftsmanship vs. automatic optimization. The battle is already over. It's cruel, but this is where we are in 2026.
But don't worry. You don't need talent or English skills to bridge this gap. You just need to know the "patterns." That's it.
In this article, I've packed all the prompt techniques and know-how I've introduced on Threads. From authentic patterns derived from research papers to tricks that exploit AI's internal structure, MCP extensions, and automation with Claude Code and Codex—about 40 items organized by chapter. All come with "copy-paste OK" prompts. You can try them as you read.
By the time you finish reading, you will fall into one of two groups: those who take these 40 tools and turn them into weapons, or those who continue to groan while writing prompts by hand. Which side you stand on depends on whether you read this to the end.
I repeat: it's OK to copy and sell. So, take it home without hesitation. Let's get started.
Chapter 1: Passing the Thought Process — Self-Verification, Self-Scoring, and Branching Thought
The first chapter is about the pattern of "passing the way of thinking itself to the AI." Japanese prompts often stop at "assigning a role and writing politely," but global pros pass "how to think." The five introduced here are all authentic patterns from research papers.
- Chain-of-Verification (CoVe) — Making AI interrogate its own answers
This is a pattern to crush hallucinations (factual errors). Normal prompts make the AI answer in one go, but CoVe is built on the idea of "making the AI verify its own answer." You have it produce a draft, convert the risks in that answer into verification questions, answer each question based on evidence, and finally present a final version with contradictions fixed. This entire flow is completed in one prompt.
When used for research or article tasks, the output becomes something else entirely. Factual errors decrease visibly, making it effective for work requiring high reliability.
Prompt for Copy-Paste
For the following theme, please complete the following steps in one response: (1) First, provide a draft answer. (2) Convert the risks of factual errors in that answer into 5 verification questions. (3) Answer each question based on evidence. (4) Present a final version with contradictions corrected. Theme: [ ]
- Self-Refine — Playing the roles of Creator, Critic, and Reviser
This pattern makes the AI play three roles in order within a single prompt. First, have it write a first draft as the creator. Next, have it score its own work as a harsh editor. Finally, have it write a finished version as the reviser based on that score.
The key is to specify concrete evaluation criteria. By specifying five criteria like "persuasiveness, uniqueness, logic, readability, and omissions," the scoring won't be lenient, and the accuracy of the revision will improve. You can feel the quality jump a level with just one prompt.
Prompt for Copy-Paste
For the following topic, please perform all three steps in one response: (1) Write a first draft. (2) As a harsh editor, score it on 5 criteria: persuasiveness, uniqueness, logic, readability, and omissions. (3) Write a revised version based on the scoring. Topic: [ ]
- Tree of Thoughts (ToT) — Branching out the answers
Instead of producing an answer in a straight line, this pattern develops multiple approaches as "branches" and lets the AI choose the best one. Think of it as taking the human process of thinking "There's Plan A, B, and C..." and putting it directly into a prompt.
This changes the dimension of brainstorming and strategic planning. Because it doesn't commit to one direction, unexpected angles emerge.
Prompt for Copy-Paste
For the following topic, please respond using these steps: (1) Develop 3 different approaches as branches. (2) Evaluate the strengths and weaknesses of each branch. (3) Select the most promising branch and provide a final answer delving deep into only that branch. Topic: [ ]
- Skeleton-of-Thought (SoT) — Framework first, flesh later
This is a thinking method where you first list the "skeleton" at the heading level and then flesh out each part. When writing long texts, if you write from the beginning, the latter half often loses steam or the structure collapses. If you solidify the skeleton first, that collapse is less likely to happen.
This is transformative for long article creation or proposal writing. Since the "blueprint" of the skeleton exists first, you won't get lost during the fleshing-out process.
Prompt for Copy-Paste
For the following theme, please follow these steps: (1) List 5 article heading skeletons in parallel. (2) Flesh out each heading as an independent separate task. (3) Finally, integrate everything into one cohesive piece. Theme: [ ]
- Meta-Prompting — Letting the AI evolve the prompt itself
This is a bit advanced. Instead of improving the content of the prompt, you ask the AI to "evolve this prompt itself." You have it create multiple improved versions, write the intent and "why it works" for each, and finally choose the strongest one.
Just by applying this once a month to the main prompts you use often, your prompt assets will grow on their own. Think of it as a system to keep your stored prompts from becoming obsolete.
Prompt for Copy-Paste
Please create 5 improved versions of the following prompt with significantly powered-up performance. For each version, write the "intent" and the reason "why it works." Finally, choose the single strongest version and state the reason for selection. Target Prompt: [ ]
What these five have in common is that they pass "thinking procedures" rather than "instructions." Self-verification, self-scoring, branching, skeleton parallelization, and self-evolution. Whether you can bake these into a single line determines if the AI is just a convenient tool or an "autonomous researcher."
Chapter 2: Backwards Design — Outcome-First Design
If Chapter 1 was about "how to make it think," Chapter 2 is about the mindset of deciding "what counts as a win" first. Japanese people use their brains on "what to throw in," while global pros write from the "goal." Here are five backwards-design prompts from official documentation.
- Output-First Specification — Fixing the final template first
Most people say "Write a blog post." But this makes the output inconsistent. In backwards design, you build the final output template first and have the AI fill in the blanks. You decide the frame first: how many characters for the title, what to include in the intro, how many headings in the body... and so on.
Output variance drops sharply, and quality stabilizes. This stability is especially effective when writing many pieces on the same theme.
Prompt for Copy-Paste
Please perfectly fill in the following template. Title: [Within 40 chars, include numbers] / Intro: [3 reader pain points, 1 sentence each] / Body: [3 H2 headings + 300 chars each] / Conclusion: [1 action proposal] / CTA: [Within 15 chars]. Theme: [ ]
- Prefilling — Specifying the start of the AI's response
This is a pattern where you specify the "opening sentence" of the AI's response. Claude is forced to write from that continuation, locking the direction of the output. Introductory greetings and unnecessary filler disappear, and format derailment almost vanishes.
It's subtle but effective. It prevents responses that make you want to snap, "So, what's the conclusion?"
Prompt for Copy-Paste
Your response must start with the following sentence: "Below, I will provide a structured response based on the requirements. First, the most important point is..."
- Negative Constraints — Listing a concrete "Don't Do" list
If you write "don't do this" vaguely, the AI won't follow it. Fuzzy prohibitions like "make it natural" are hard to enforce. But if you list them concretely, it follows. Prohibit honorifics, prohibit opening greetings, prohibit specific phrases... list them in bullet points.
Generic AI-like patterns disappear significantly. This is a staple when you want to remove the "obviously generated" feel from text.
Prompt for Copy-Paste
Please create the following. However, strictly adhere to these prohibitions: (1) No honorifics. (2) No repeated use of 3-character compound words. (3) No expressions like "regarding..." or "it is important to...". (4) No opening greetings. (5) No simple bulleted lists. If violated, rewrite everything. Target: [ ]
- XML Structured Tagging — Separating information with tags
Separating information with tags improves the AI's reading accuracy. Goal, background, constraints, reference examples, output format. Instead of throwing these in as a lump of text, partition them with tags. Pros don't throw text; they throw blueprints.
Prompt for Copy-Paste
I will structure the prompt as follows. Please respond according to the content within each tag. <goal>Goal to achieve</goal> <context>Background information</context> <constraints>Prohibitions</constraints> <examples>Reference examples</examples> <output_format>Output format frame</output_format>
- Persona Stack — Layering roles in 3 levels
Most people stop at "You are a copywriter." Pros layer roles in three levels, not just one. A writer role, an editor role, and the target reader role. By giving one AI these three personas simultaneously, you run the cycle of writing → editing → re-correcting from the reader's perspective in one go.
Because multiple perspectives run simultaneously, persuasiveness increases. However, there are caveats to this "persona" approach, which we will cover in Chapter 10.
Prompt for Copy-Paste
Please assume the following 3 personas simultaneously: (1) Top Copywriter (Writer). (2) Harsh Editor-in-Chief (Editor). (3) Target Reader, a 30s office worker (Recipient). Execute writing → editing → re-correction from the reader's perspective all in one response. Theme: [ ]
The essence of backwards design is simple. If the first line doesn't have the "goal," "prohibitions," and "format frame," the AI is already lost. It's the difference between writing from the entrance or the exit.
Chapter 3: Running AI as a "Legion" — Multi-Agent Operation
From here, the mindset changes. Instead of exhausting one AI, you assign multiple roles to AI and operate them as a "legion." Fighting with just one is like going to a battlefield with one weapon. Global pros have entered the phase of commanding from a general staff headquarters.
Note that you don't need to set up multiple AI accounts for these patterns. You can reproduce them just by "switching roles" within a single chat.
- Routing Pattern — Placing a sorter
This is a military-style organization where a "classifier" AI receives the input first and routes it to the appropriate "specialist." Most people throw everything at one AI and end up with a mediocre answer. Pros place a receptionist to pass the task to an expert.
Accuracy improves when you pivot to specialization rather than one AI pretending to be omnipotent.
Prompt for Copy-Paste
You are the "Routing Coordinator." Read the following request and classify it into: (1) Research, (2) Writing, (3) Analysis, or (4) Code. Create a 5-line system prompt optimized for that classification and execute the request again using it. Request: [ ]
- Parallelization — Solving the same question in parallel for a majority vote
This pattern involves solving the same task from multiple perspectives in parallel and determining the conclusion by majority vote. Instead of asking one expert, you have five experts answer with different approaches and take the conclusion supported by the most. That's the idea.
Hallucinations decrease. When you want to leave important decisions to AI, the sense of security is completely different.
Prompt for Copy-Paste
For the following question, first provide one answer each from the perspectives of 5 independent experts using different approaches. Next, compare the 5 plans, take the conclusion supported by the most perspectives as the final answer, and state the reason for selection. Question: [ ]
- Evaluator-Optimizer — Completely separating the creator and evaluator
Self-evaluation done by one person inevitably becomes lenient. If you separate the actor and the judge, the judgment becomes harsh. You reproduce this within one prompt. Create the best answer as the creator, completely switch personas to a harsh evaluator to score it, then return to the creator to make the final version.
Similar to Self-Refine in Chapter 1, but the point here is to explicitly state "completely switch personas."
Prompt for Copy-Paste
For the following topic, first create the best answer as the "Creator AI." Next, completely switch personas and, as the "Harsh Evaluator AI," score that answer out of 100 and list 5 reasons for point deductions. Finally, return to the Creator role and create a final version based on the evaluation. Topic: [ ]
- Multi-Agent Debate — Debating with Proponents, Opponents, and a Moderator
This is a pattern where multiple AIs debate before an integrator reaches a conclusion. Proponents and opponents clash, and a neutral moderator summarizes their points. Extremes and thought-stopping disappear, resulting in a balanced conclusion.
This shows particular strength in "questions where it's hard to decide," such as strategic planning or decision-making.
Prompt for Copy-Paste
For the following theme, please act as: (1) Proponent AI, (2) Opponent AI, and (3) Neutral Moderator AI. Have (1) and (2) debate for 3 rounds each, then have (3) integrate their points to provide a final conclusion and reasoning. Theme: [ ]
- Self-Verifying Output — Interrogating yourself before outputting
This is a pattern where the creator "interrogates" their own output as a different persona before final submission. Since it's hard to see flaws in your own writing, you force a change in perspective. You switch through three personas—global competitor pro, target reader, and harsh boss—and have each list problems.
Recent AI models are evolving toward "reporting after verifying output themselves." Think of this as a pattern that anticipates that movement from the prompt side.
Prompt for Copy-Paste
Please re-read the following output as completely different personas. Switch through the roles of (1) Global Competitor Pro, (2) Target Reader, and (3) Harsh Boss in order, list 3 problems from each perspective, and finally present one strongest improved version. Output: [ ]
The essence of legion operation is whether you can stand on the premise that "AI is not something to be used as a single unit." Assign roles and command. That alone changes the results coming from the same AI.
Chapter 4: Designing Context as an "Environment" — 4-Layer Thinking
Up to now, we've talked about the "content of the prompt." But in the world of global AI designers, the prompt is treated as the "bottom layer" of a much larger structure: Prompt → Context → Intent → Specification. Most people are still stuck at the first layer. This chapter is about five ways to climb to the layers above.
- Bookend Placement — Placing important constraints at both the beginning and end
When you pass long text to an AI, information placed in the middle tends to lose attention. This is the "middle is a blind spot" phenomenon. Therefore, always place important constraints in two places: the beginning and the end. Imagine sandwiching important things at both ends like "bookends."
This is effective for people who write long prompts. A common failure is having a long instruction where the constraint is only written once in the middle.
Prompt for Copy-Paste
In the following request, please restate important constraints in two places—the "beginning" and the "end" of the prompt—before processing. Structure it on the premise that information in the middle of the context is prone to distraction. Request: [ ]
- Goldilocks Altitude — The "just right" altitude for system prompts
System prompts have an optimal "altitude." Too low (binding with detailed if-else logic) causes rigidity; too high (abstract philosophy only) decides nothing. Aim for the altitude in between. This is the "Goldilocks Zone" concept applied to prompts.
The trick is to design in three layers: principles that never change, frameworks for situational judgment, and freedom for individual tasks. Writing with these three in mind avoids both rigidity and abstraction.
Prompt for Copy-Paste
I will fix instructions to you in the following 3 layers: (1) Invariant Principles (Why/Never change). (2) Framework for Situational Judgment (When/Branching rules by case). (3) Freedom for Individual Tasks (What/Leave to your judgment). Design at a "just right" altitude, avoiding both rigidity and abstraction.
- Just-In-Time Context Injection — Passing only what's needed when it's needed
Just because the context window is large doesn't mean stuffing all materials in is good; it can actually degrade accuracy. Pros first pass only the "Table of Contents, Summary, and Index" and have the AI retrieve necessary chapters as needed. Imagine not stacking all books from the library on your desk, but going to borrow only the one book you need.
The correct way to handle large materials is "dynamic calling," not "stuffing everything."
Prompt for Copy-Paste
I will not input the following massive material all at once. In the first stage, please read only the "Table of Contents," "100-character summary of each chapter," and "Index." If there is a chapter where details are needed, explicitly request it from me before retrieving it, and add only that chapter to the context for work.
- Intent Encoding — Articulating judgment criteria first
Instead of explaining your organization's or your own "values, priorities, and trade-off judgment criteria" from scratch every time, articulate them once and pass them. Just by placing this at the beginning, the AI starts moving as "your agent." It becomes an AI that doesn't just wait for instructions but can lean in the right direction on its own when lost.
Prompt for Copy-Paste
As a premise for the following work, I will articulate my judgment criteria: (1) Priority [A > B > C]. (2) Absolute NG [X, Y, Z]. (3) Default judgment in ambiguous cases [D]. When lost in judgment, always return to these criteria before deciding.
- Specification Layer — Standing on the side of creating "Specifications"
The top of the 4 layers is this "Specification." Fix quality standards and business rules as structured text (specifications) and make that the starting point of the context every time. From a person who writes prompts to a person who creates specifications. The moment you stand here, the reproducibility of work jumps instantly.
Prompt for Copy-Paste
Future work will refer to the following "Specification (Markdown structured format)" as the starting point for every session. If judgment outside the specification is required, do not guess; always check with me. [Paste your specification here]
The idea of this chapter is the transition from "competing with a single line of prompt" to "designing the entire context as one environment." It's the difference between fighting on one layer or four.
Chapter 5: Aligning with AI's Internal Mechanics — Structural Design for KV Cache
This chapter is a bit geeky. But knowing this changes the AI's "speed," "cost," and "how usage limits decrease."
Inside the AI, a mechanism called "KV Cache" is running. To put it very simply, the AI keeps processed content internally, and when the same content comes again, it can reuse it. Conversely, if you use it in a way where reuse doesn't work, it recalculates from scratch every time.
"Hitting usage limits quickly," "responses getting slower as the conversation gets longer," or "API bills being higher than expected." The cause of these troubles is often not the content of the prompt, but the "placement" being out of sync with the AI's internal mechanics.
- Stable Prefix First — Fixing static things at the beginning
Cache works "only for the part that matches perfectly from the beginning." Therefore, fix things that don't change (premises, reference materials, rules) at the beginning, and place things that change every time (today's question) at the end. Just by putting the same premise at the head every time, the cache starts working.
Prompt for Copy-Paste
I will fix the premises for subsequent work: (1) My industry = [A]. (2) Target = [B]. (3) Prohibitions = [C]. (4) Output format = [D]. Please re-declare this at the beginning of every session before entering the main topic.
- Anchor Document Pattern — Throwing materials only once at the start
Are you re-pasting large reference materials or guidelines every time? That's a waste. Throw the materials only once at the beginning as an "anchor," and make subsequent questions just refer to those materials.
Prompt for Copy-Paste
I will make the following materials the anchor for this session: [Bulk input reference materials here]. Henceforth, please answer all questions I ask by referring to these materials. Re-presentation of materials is unnecessary.
- Session Continuity — Continuing related work in one thread
Every time you open a new chat, the AI rebuilds its internal cache from scratch. Opening 10 new chats a day vs. continuing in one long session once a day changes how usage limits decrease and the output quality. Frequent new chats are part of the reason for "hitting limits quickly."
Operational Guideline:
Always continue work on the same theme in one thread. If the thread gets too long, create a "summary so far" at the beginning and continue using that as an anchor.
- Differential Edit Pattern — Instructing only the differences when fixing
When you want to improve output, re-posting the entire text is NG. Re-posting everything clears the precious cache and recalculates from the beginning. Instruct only the differences: "Only this part," "Change this part to that."
Prompt for Copy-Paste
When I want to fix previous output, I will not re-post the entire text. I will instruct only the differences, like "Only the part [ ]" or "Change [ ] to [ ]." I will not re-present premises or reference materials at all.
- Cache-Aware Sub-agent Design — Aligning the beginnings of sub-agents
Even when doing legion operations like in Chapter 3, aligning the "beginning part" (role definition, premises, rules) of each agent's system prompt makes the cache easier to hit. Change only the task content individually at the end. This alone improves the efficiency of legion operations.
Prompt for Copy-Paste
When performing multi-agent operations, perfectly unify the beginning part (role definition, premises, rules) of all agents' system prompts. Change only the task content individually at the end.
The essence of this chapter is "competing with structure, not just content." Where you place things matters as much as what you write.
Chapter 6: Creating the "Outside" of the Prompt — Harnesses and Agents
Entering 2026, global AI developers stopped competing over the "content of the prompt" and started designing the "outside of the prompt." That outside is called a "harness." First, let's organize the big picture.
What is a "Harness Agent" anyway?
The ChatGPT or Claude you normally use is actually not an "AI agent." It's just a "brain unit." An AI agent refers to the state where parts are attached to that brain to turn it into a "self-running machine."
An agent consists roughly of the following elements:
- Model (The Body): The agent's "intelligence." The LLM itself. With just a brain, it can make judgments but cannot act.
- Harness: The set of instructions (system prompt) and guardrails (what not to do) given to the model. For example, a safety valve like "Always get human confirmation for payments exceeding a certain amount." Think of it as the layer that determines the agent's "personality" and "judgment criteria."
- Tools (Hands and Feet): The agent's interface to touch the real world, such as sending emails, operating calendars, reading/writing files, and web searches. Without tools, AI can read a receipt but cannot submit it for expense reimbursement.
- Environment: "Where" the agent moves. Even with the same AI, what it can do changes greatly depending on the environment it moves in.
- Agent Loop (Self-Running Cycle): All these work together to keep running a loop of Plan → Act → Observe Result → Adjust → Repeat. It stops if human confirmation is needed. The decisive difference between a one-question-one-answer chatbot and a self-running agent is here.
These five elements combined are the "complete form of an agent." From here, I will introduce five techniques for highly designing that outside = harness.
- Execution Loop — Incorporating a cycle of Observation, Thought, Self-Criticism, and Action
This is the heart of the harness. When executing a task, explicitly have it follow the cycle of "Observation → Thought → Self-Criticism → Action" at every step. It moves on a different dimension from a one-shot "request."
Prompt for Copy-Paste
When executing the following task, please always write the following 4 stages in order at each step before proceeding: (1) Observation: Describe the current state in 3 lines. (2) Thought: The next necessary move and reason. (3) Self-Criticism: One blind spot of that move. (4) Action: The final action after correction. Task: [ ]
- Context Compaction — Compressing long conversations in stages
As conversations get longer, the context breaks down. To prevent this, forcibly compress past interactions into a fixed format every certain number of rounds. This is essential design to keep context from "rotting" in long-form tasks.
Prompt for Copy-Paste
Henceforth, every time the conversation exceeds 10 rounds, forcibly compress past interactions into the format of "3 confirmed facts + 2 pending tasks + 1 next most important action," and execute new instructions starting from that compressed version.
- Playbook Memory — Accumulating reusable "patterns"
This is a design to accumulate strategies as reusable "patterns (playbooks)" instead of having them generated from scratch every time. Each interaction becomes an "asset."
Prompt for Copy-Paste
After executing the following task, always output "3 general rules from this time" in a Markdown bulleted list. From next time, I will present them at the beginning, so please read them as a playbook before working.
- Self-Modification Loop — Letting AI rewrite its own instruction manual
This is a pattern where the agent itself writes a "self-instruction template" to do better next time. You can create a state where the AI keeps updating its own instruction manual every time it's used.
Prompt for Copy-Paste
After executing the following task, please output a "self-instruction template (improved version)" for yourself to execute with better accuracy next time. Also, state the intent and aim of the improvement. I will use that template from next time.
- Auto-Harness Optimization — Iteratively improving the entire mechanism
Similar to Meta-Prompting in Chapter 1, but the target is the "prompt structure itself." Just by running this once a week, your prompt assets will evolve with compound interest.
Prompt for Copy-Paste
Please provide 5 concrete plans to improve my current prompt structure by one level. Write the aim, expected improvement effect, and expected risk for each plan, and finally choose the strongest plan, stating the reason for adoption and a next usage example.
Polishing prompt content line by line is important. But there is a much larger world "outside" of that. Whether you can have the perspective of a harness changes how you use AI by one level.
Chapter 7: 5 Secret Prompts No One Knows
This chapter collects niche patterns that are rarely found in standard prompt books. Some are from research, others apply cognitive science. These are angles rarely shared in Japan.
- Question Reframing — Having it reframe into a neutral question before answering
AI has a habit of "pandering to the user (over-conforming)." It anticipates whether to agree or disagree with your statement and returns a sycophantic answer. As a way to suppress this, it's reported that having it reframe your statement into a "neutral question" before answering is more effective than directly commanding "don't pander."
Try using this when you want an opinion but only get "It's exactly as you say."
Prompt for Copy-Paste
Before answering my following statement, please always first reframe my statement into a "neutral question format." After declining to express agreement or disagreement with the original statement, please answer logically. Statement: [ ]
- Verbalized Sampling — Having it output multiple plans with probabilities
AI tends to return similar answers every time you ask the same question (a phenomenon where output diversity collapses). The countermeasure is a pattern where you have it output multiple answer plans with the "probability they should be chosen." Choosing one plan from the probability distribution can draw out ideas that don't emerge with normal prompts.
Prompt for Copy-Paste
For the following topic, please generate 5 different answer plans with their respective "probability of being chosen (%)." The total probability must be 100%. Finally, choose and present one plan from that distribution according to the probability. Topic: [ ]
- Counterfactual Anchoring — Having it output the opposite answer first
This is a pattern applying the "anchoring effect" from cognitive psychology. By having the AI first output the "completely opposite answer" to the one it intuitively wants to give, you break the inertia toward a mediocre solution. A predictable answer suddenly becomes deep.
Note: "Counterfactual Anchoring" is my coined term. The underlying principle (anchoring effect) is real cognitive science, but this specific pattern is an application.
Prompt for Copy-Paste
Before answering the following question, first generate one "completely opposite answer" to the one you intuitively intend to give. Write 5 grounds for why that opposite plan might be correct. Finally, provide the final answer after going through that verification. Question: [ ]
- Self Pre-Mortem — Listing causes of failure first
"Pre-Mortem" is a famous thinking method where you think "If this project failed, what would be the cause?" before starting. We apply this to AI task execution. Have it list failure patterns before execution and have it execute in a way that avoids them. Effective for long-form tasks.
Note: The original Pre-Mortem method is real, but applying it to AI is my own usage.
Prompt for Copy-Paste
Before executing the following task, first list "5 assumed causes if this task ended in the lowest quality." Write a one-line avoidance measure for each cause, and execute the task while strictly following all those avoidance measures. Task: [ ]
- Calibrated Confidence Prompting — Making it state confidence levels
AI's biggest weakness is "saying things with full confidence even when wrong." To stop this, use a pattern where it must accompany each claim with a "confidence level (0-100%)." Hallucinations become visible, making it much easier to judge the reliability of information.
Prompt for Copy-Paste
When answering the following question, please accompany each claim with a "confidence level of 0-100%." Label less than 50% as "Guess" and only 70% or more as "Fact." Also, provide a one-line basis for each confidence level. Question: [ ]
Whether you study them as standard techniques or operate them as secret tricks from papers and principles, most sources are available for free. The gap widens starting with those who notice.
Chapter 8: Giving AI "Hands and Feet" — MCPs You Should Install
Up to now, we've talked about prompts. This chapter is about giving AI "hands and feet."
MCP (Model Context Protocol) is an open common standard for connecting AI to external services and data. Simply put, it's a mechanism to add "windows to touch the real world" to AI. Installing this turns Claude from a chatbot into an "agent with hands and feet."
Since MCP is an open standard, once you set it up in Claude Code or Codex, the same environment can be carried over. Here, I've selected four that are truly useful for content creators.
- Supadata MCP — Extracting transcripts from videos in one shot
This is an MCP that can extract transcripts from YouTube, TikTok, Instagram, and X (formerly Twitter) videos in one shot. Overseas content research, competitor video analysis, and trend grasping become much faster. Even for videos without subtitles, it can be covered by the automatic transcription function.
Usage Example:
Extract the transcript from the following YouTube/TikTok/X video URL and summarize it into 5 key points. URL: [ ]
You can graduate from the time-consuming task of watching overseas viral videos to the end.
- Firecrawl MCP — Converting websites into clean Markdown
This is an MCP that converts any website into clean Markdown that AI can easily read. It can even process pages rendered with JavaScript. It becomes faster to have AI read competitor articles, LPs, or news sites to output structural analysis or improvement proposals.
Usage Example:
Markdown-ize the following URL with Firecrawl and analyze the structure and weaknesses of the appeal in that article. URL: [ ]
People who copy and paste by hand every time can get their time back here.
- Google Knowledge Graph MCP — Direct access to entity information
This is an MCP where AI can directly access the source data of the "Information Panel" that appears on the right side of Google searches. You can pull structured data on real people, places, organizations, and concepts. Accuracy changes in fact-checking and verification of person/organization information. It becomes a lifeline for information publishers.
Usage Example:
Get information on "[Person/Organization Name]" with Google Knowledge Graph, and then check the factual relationships in my article.
- Memory MCP — Giving AI permanent memory
AI has a weakness where "conversation history is reset from scratch every time." Memory MCP solves this. Project decisions, your preferences, and rules learned in past interactions remain even after the session ends.
Usage Example:
Register the following preferences, premises, and ongoing project information as permanent memory in Memory MCP. From next time, always refer to this content first before working.
The more you re-paste the same premises in every session, the more you'll feel the effect. Claude grows into a "partner who understands you" as sessions repeat.
With these four, you have both "Input (Video, Web, Entity Information)" and "Memory." Accumulate the information gathered with the first three into the AI using Memory MCP. The effect when the four work together is greater than using them individually.
Chapter 9: Automating with Claude Code & Codex — 5 Points and Pitfalls
If you are "somehow running" automation with Claude Code or Codex, you are losing out quite a bit. This chapter introduces five points for getting results with automation, along with their respective pitfalls.
- Always insert Plan Mode
The lifeline of automation is not letting it execute suddenly. In Plan Mode, have it concretize the filenames, function names, and sequence of steps to be edited before you approve. Always insert this.
Pitfall: Skipping Plan Mode and jumping to parallel execution. Parallelization without a planning gate just mass-produces wrong deliverables at high speed. Speed is only valuable when the direction is correct.
- Durable rules in config files, per-time instructions in prompts
It's a typical beginner mistake to stuff "permanent rules you want followed every time" into every prompt. Write durable rules in configuration files (AGENTS.md for Codex, CLAUDE.md for Claude Code), and put only "instructions for this time" in the prompt.
Pitfall: Making the configuration file too large. If it exceeds the size limit, content will be cut. If the file gets large, the standard practice is to split it by directory hierarchy.
- Sub-agents should be "Specialized + Restricted Permissions"
Assign "1 function = 1 specialized role" to sub-agents and give them only the minimum necessary tools.
Pitfall: Sub-agents inheriting all of the parent's tool permissions by default. Since they start with full permissions, it can lead to accidents if you don't explicitly restrict them. Also, sub-agents consume more tokens as they run multiple models and tools. Since parallelization = increased cost is structurally unavoidable, narrow it down to tasks worth running.
- "Don't overstuff" is the right answer for MCP
I introduced MCP in Chapter 8, but overdoing it is forbidden. The more MCPs you add, the more the context of each message swells, squeezing the usage limit. Deactivate MCPs you don't use. This is basic.
Pitfall: MCP-izing everything and plugging in 10 or 20. In addition to squeezing context, security risks also increase. It's safe to narrow it down to 3-5 that you truly use every day.
- Package repetitive work as "Skills"
For workflows you use repeatedly, don't copy-paste the prompt each time; package it as a Skill. Both Claude Code and Codex have mechanisms to summarize repetitive work into definition files and call them consistently.
Pitfall: Operating by copy-pasting prompts every time without Skill-izing. Variance accumulates, and the mental load keeps increasing. Any workflow with 2-3 concrete usage examples is ready to be Skill-ized. Writing them on the premise of using the same Skill in both Claude Code and Codex ensures no cost even if you switch tools.
The essence of automation is the difference between "dumping everything" on AI or "systematizing" it. Insert Plan Mode, organize config files, restrict permissions, minimize MCP, and Skill-ize repetitions. Set up these five mechanisms first before running. That alone determines whether automation becomes an "acceleration" or a "runaway."
Chapter 10: It's Obsolete — Prompt Habits You Should Stop
The final chapter is about subtraction, not addition. I'll cover three habits that have long been considered "correct" in Japan but are being reconsidered in the latest research and specifications.
- Attaching "You are an expert in..." to everything
Prompts that give a persona are a classic staple. However, recent research points out they are "not omnipotent." While expert personas are effective for safety and moderation tasks, reports suggest they tend to actually lower accuracy in factual recognition and reasoning.
In other words, a persona is not a "universal template" but a "pinpoint technique to use by choosing the application." It's okay to attach it to safety judgment, ethical judgment, or moderation. But for fact-finding, analysis, code generation, or reasoning tasks, don't attach it needlessly. Just by being conscious of this distinction, accuracy goes up a level.
- Mechanically attaching "Think step by step"
"Think step by step" and "Let's think step by step" have also long been treated as the strongest prompts. However, recent models are evolving toward the model itself judging when and how much to think. The technical necessity of writing "think" every time is fading.
Future operation should focus on articulating the task's purpose, constraints, and expected format rather than mechanically attaching thinking instructions. And for complex tasks requiring reasoning, turn on the model's own thinking function and leave the judgment to the AI. This is more compatible with the new generation of models.
- Writing prompts by hand and fine-tuning by intuition
This might be the biggest point. Many Japanese people write prompts by hand as "works of art" and fine-tune them by word choice and intuition. Meanwhile, global pros treat prompts as "code." They design, version control, test, and optimize while scoring with evaluation criteria (eval).
Why is it so different? The reason is simple. Hand-written intuition cannot detect the "10% of cases where a prompt that works for 90% of inputs fails catastrophically." Therefore, they first create a scorecard for "what counts as a correct answer" and measure the prompt against it.
The first step toward "eval thinking" you can do today:
(1) Run the same prompt 5 times and observe the variance in output. (2) Write down 3 conditions common to good outputs (this is a simple eval). (3) Create multiple versions of the prompt at the word level and compare them until they satisfy those 3 conditions.
By the way, even with words that mean the same thing, the output changes. The AI's reaction is subtly different between "Calculate" and "Compute." Hand-written intuition cannot control this word-level difference. That's why there's value in switching your brain to "measure and improve."
I've introduced many techniques to add in Chapters 1 through 9. But those who grow also decide what to "stop."
Conclusion — Prompts Have Become "Blueprints for Thought"
So far, I've introduced about 40 techniques and know-how in a whirlwind. Finally, I'll summarize what I wanted to convey in this article into one thing.
Prompts are no longer "instructions." They are "blueprints for thought."
In Chapter 1, we passed the way of thinking itself. In Chapter 2, we worked backwards from the exit. In Chapter 3, we operated AI as a legion. In Chapters 4 through 6, we designed the outside of the prompt = context, internal structure, and harness. In Chapter 7, we used secret tricks; in Chapter 8, we gave AI hands and feet with MCP; in Chapter 9, we set up automation mechanisms; and in Chapter 10, we let go of old habits.
What they all have in common is that they write "how to make it think and move," not "what to output." Commands vs. Blueprints. That difference determines whether AI remains "just a convenient tool" or turns into a "self-running partner."
The techniques introduced here require neither talent nor English skills. All you need is to try writing one line of your prompt from today with a slightly different structure. Copy and paste one pattern that caught your eye and try it. That will be your first step.
Thank you for reading to the end. I hope this article serves as a catalyst to raise your relationship with AI by one level.
Consult about AI here 👇
