Ichiro Furutachi Could Conquer the AI Agent EraWhy Voice Prompting Can Become a Real Advantage

Ichiro Furutachi Could Conquer the AI Agent Era: Why Uncompressed Verbalization Becomes a Competitive Advantage

I have argued this point before, but there is still a major misunderstanding about how people should give input to generative AI and AI agents. Many people still assume that the ideal prompt is short, polished, typo-free, and logically organized. For simple Q&A or narrow correction tasks, that view has some validity. If the scope is limited, the objective is clear, and the shape of the answer is mostly fixed, a compact prompt is faster and easier to handle.

But the situation changes when the work is about creating something from zero to one: imagining a new product, designing functions, tuning the emotional temperature of UI and UX, shaping a brand direction, or articulating the philosophy behind a service. At that stage, what matters most is not a neat command sentence. What matters is information that is still close to the raw thinking itself. There are moments when uncompressed input, including hesitation, discomfort, exceptions, anxiety, non-negotiable values, and even apparently contradictory wishes, gets closer to the real goal than a beautifully summarized prompt ever can.

This is not merely a matter of prompt-writing technique. It is a deeper question about how we understand the relationship between human thought and AI. In particular, methods like voice input, where people speak while thinking, may be seriously underestimated in the AI era. Recent research and practical experience suggest that this deserves much more attention.

---

There are cases where uncompressed context beats tidy short writing

In the early days of generative AI, people repeatedly advised users to write prompts as briefly, structurally, and clearly as possible. Use Markdown bullets. State the goal, constraints, and output format. Remove ambiguity. That advice was not wrong. When models had weaker contextual understanding, a large part of the burden really did belong on the human side.

However, recent models, especially those that can handle long contexts, multi-step reasoning, and agentic planning with tool use, reward a slightly different quality of input. What matters is no longer the visual beauty of the request. What matters is whether the information necessary for decision-making is still present.

Short prompts often contain self-compression. People try to make things easier to understand, and in the process they cut away background circumstances, edge cases, doubts, emotional resistance, and hidden concerns. The sentence becomes cleaner, but the design intent becomes thinner. Because the person still has the full picture in their own head, they feel as if they have communicated enough. In reality, the AI only receives the thin cross-section left after summarization.

By contrast, a long input that wanders a little, contains typos, and keeps revising itself with phrases like “no, not that” or “actually, this is what I want” often preserves information that never appears in a formal specification. It keeps the balancing judgments that define real product design: “I want it fast, but not cheap-looking.” “I want it kind, but not preachy.” “I want it powerful, but not scary to first-time users.” “I want future extensibility, but not immediate complexity.” Those tensions are often the substance of the design itself.

So the real comparison is not between one thousand polished words and ten thousand messy ones. The real comparison is between how much information gets lost through self-summary. Long writing is not inherently stronger. What is stronger is context that has not yet been prematurely deleted.

Voice input can prevent thought from being compressed too early

This is where voice input becomes extremely important. People tend to shrink their thinking to match typing speed. When they speak instead, that compression happens less easily. Natural speaking speed is often estimated at roughly 120 to 200 words per minute. In a 2016 smartphone input study by researchers associated with Stanford University and the University of Washington, English speech input reached 153 words per minute while keyboard input reached 52 words per minute. That is about a 2.9 times difference. A similar 2.9 times speed advantage was reported for Chinese.

That gap is not just a matter of convenience. It changes thought itself. When input is slow, people unconsciously start omitting things because writing them down feels tiresome. Small concerns, half-formed discomfort, exception handling, and subtle temperature differences that feel important but are still difficult to articulate often disappear for no reason more sophisticated than “it is too much work to type.” The result is that the AI receives a version of the idea that is far cleaner and far thinner than what the person was actually thinking.

Voice input reduces that loss. For people whose thoughts become clearer while they speak, the effect can be decisive. There are people for whom talking is not the final expression of thought but the process of design itself. Speaking is how they search a design space, test hypotheses, and discover structure. For them, voice input is not just an input method. It is an interface that connects their thinking speed directly to AI.

For convenience, let us call such people “talking engineers.” They do not sit silently, summarize the key points, and then write. They explore the design space by speaking and can hand that exploratory process directly to AI. In conventional workplaces, such people may have been judged as too verbose, too scattered, or too prone to talking to themselves. In the AI era, however, that apparently messy process can become extraordinarily valuable primary information.

But speaking is not always superior

This argument should not be generalized too far. Voice input and oral thinking have weaknesses. A 2015 study on creative ideation found that idea fluency was higher in writing conditions than in speaking conditions. The researchers pointed out that speaking tends to place a heavier load on working memory and can increase cognitive cost. In other words, not everyone thinks better by talking.

That point matters. Voice is fast, but speed and depth are not the same thing. For some people, speaking pushes thought outward in a productive way. For others, the cognitive burden of maintaining coherence while speaking narrows their imagination instead of expanding it. That is why the simplistic conclusion “everyone should switch to voice input” does not hold.

Yet the reverse is also true. In many workplaces, people who can think clearly in writing are treated as the canonical form of intelligence, while people who deepen thought through speech are underestimated. In collaboration with AI, that balance is likely to change. Especially in the stage where vague ideas must be turned into concrete requirements, the ability to think out loud may become not merely comparable to writing skill, but in some cases more powerful.

Even the old weakness of voice input, recognition errors, has to be reevaluated. Research on medical documentation has shown that speech recognition can improve writing speed, document length, and user satisfaction, while still creating a burden of correction. A 2018 study in JAMA Network Open found 7.4 errors per 100 words in initial speech-recognized documentation. But in today’s generative AI environment, those rough transcripts do not always need to be corrected by hand first. AI can often restructure, summarize, and clean them up. The old drawback of messiness and recognition noise is no longer as fatal as it once was.

What matters to AI is not length but relevant specificity

At this point, it is important not to fall into blind worship of long prompts. The point is not that longer is always better or that more text automatically means more value. What AI truly benefits from is sufficient relevant specificity, not raw character count.

Recent research suggests that vague, low-information prompts, in other words underspecified instructions, make model performance more unstable. A 2026 study reported that prompts containing only minimal guidance showed greater variance and lower stability than prompts with more concrete conditions. Another 2025 study found that in reasoning tasks, detailed instructions tended to outperform vague ones. That makes intuitive sense: models infer what to prioritize from the prompt itself.

This tendency is especially clear in software development. A 2026 code-generation study organized effective prompt improvements into categories such as input and output format, preconditions, postconditions, exception handling, dependency disclosure, and ambiguity reduction. Among fifty practitioners surveyed, explicit I/O specification was seen as particularly useful. In other words, the value for AI lies not in elegant phrasing but in whether the conditions needed for good decisions are actually present.

From this perspective, a polished thousand-word prompt loses not because it is short, but because the act of shortening often deletes the most important conditions. A messy ten-thousand-word prompt wins not because it is long, but because it preserves design constraints and philosophy that would otherwise be dropped.

Long context still has hard limits and real risks

At the same time, long context has obvious limits. A 2025 long-context study reported cases in which model performance fell by 13.9 percent to 85 percent as input length increased, even when the relevant information was technically available. In other words, a longer context does not automatically make a model wiser. Once irrelevant details, repetition, noise, and side branches accumulate, the model can become easier to distract.

This matters even for people who value uncompressed voice input. Raw context has value, but that does not mean it should always be used as the final execution instruction. The stronger workflow is two-stage: first, get the uncompressed material out in full; then have AI reorganize it.

That idea also aligns with recent research. A 2025 study by researchers from Microsoft, the University of Maryland, and others showed that response quality improved when another LLM rewrote low-quality user input based on real human-AI conversation logs. For information-seeking questions, rewritten prompts were judged better than original prompts in more than 70 percent of evaluations with GPT-4o. Human evaluations suggested that 74 percent of rewrites strongly preserved intent, 21 percent preserved it somewhat, and only 5 percent substantially missed it. Even reasoning-focused models gained roughly ten percentage points in win rate when a rewriting step was inserted.

That is highly suggestive. In modern AI practice, users do not need to write the perfect prompt on the first try. In fact, it can be more effective to dump out raw conversation, hesitation, side notes, detours, clarifications, and self-corrections, and then ask AI to rewrite everything into a specification that preserves the purpose, constraints, and priorities.

In zero-to-one development, over-summarizing is itself a loss

This issue becomes most visible in new product development. If you are fixing a local bug in existing code, a concise note containing reproduction steps, error logs, expected behavior, and actual behavior is often enough. But when you are launching a new service, that is nowhere near sufficient.

Zero-to-one work contains an enormous amount of information that rarely survives in a formal requirements document. What emotion should the product evoke, and in whom? How should it differ from competitors? Where should it feel premium, and where should it feel intentionally approachable? How far should automation go, and where should human space remain? Should the monetization flow be assertive or restrained? What exactly makes users anxious, and what makes them feel safe? How simple should the first experience remain even while future extensibility is being considered? These questions often contain the core of the design, yet they are hard to preserve in conventional specifications.

That is why people who can pour their actual thinking process into AI, including moments such as “that is wrong,” “this part is non-negotiable,” “this would be too much,” or “but if it is too weak it will not land,” are so strong. The crucial skill is not the ability to summarize beautifully from the start. It is the ability to pass the weightings, tensions, and hesitations of the design into AI without losing them.

Generative AI is not just a tool for polishing prose. It is also a tool for updating design proposals while considering multiple conditions at once. The better a person can feed the “constraints that are not fully verbalized yet,” the better the output tends to become. Half-clean instructions often produce polished but fundamentally misaligned results. Rougher input, if it is rich in design intent, is often easier to refine toward the true target through iteration.

Ichiro Furutachi symbolizes an elite form of verbalization in the AI era

This is why Ichiro Furutachi matters here, not simply because he is famous, but because he embodies a kind of ability that becomes extremely valuable in the AI era.

Furutachi joined TV Asahi in 1977 and established the distinctive style later called “Furutachi-bushi” through professional wrestling commentary. He later became widely known for Formula One commentary, served as the host of the NHK Kohaku Uta Gassen for three consecutive years, and then spent twelve years as the anchor of Hodo Station. He has remained active across commentary, hosting, and news, and today also serves as a visiting professor at Rikkyo University. The key point is not just the breadth of his career. It is that in every domain he demonstrated extraordinarily advanced verbalization.

His commentary is not mere explanation. It reconstructs what is happening in front of him inside the viewer’s mind, using metaphor, speed, structure, and emotional contrast. He does not simply describe visible facts. He captures heat, danger, and momentum before they are fully recognized, then renders them visible through apt metaphor. That is a different skill from simply possessing information. It is the ability to transform observed phenomena into dense language that another person can grasp.

For AI agents, that ability is enormously powerful. Agents act on goals, constraints, priorities, and evaluation criteria. In that environment, the crucial thing is the ability to verbalize the essence of the target with neither too little nor too much. Conditions like “solid but not old-fashioned,” “sharp but not aggressive,” or “kind but not over-explanatory” are difficult for many people to articulate. A speaker on Furutachi’s level may be able to transmit those conditions in multiple layers through metaphor and structure.

If someone like that mastered AI agents in earnest, they could become an extraordinarily powerful player. In the AI era, the advantage does not belong merely to people who have knowledge. It belongs to people who can verbalize the structures, images, nuances, concerns, and expectations inside their heads quickly and richly. In that sense, Furutachi is a vivid symbol of input power in the age of AI agents.

Companies should not overlook talking engineers

This argument is not only about personal productivity. It also affects organizational design. Many companies are now pushing for a return to the office. Face-to-face conversation, teaching, and immediate coordination certainly have value, and that judgment is understandable. But if companies apply that model uniformly to everyone, they may crush part of the talent that will thrive in the AI era.

Talking-engineer types sometimes do their best work in quiet environments. They externalize in-progress design through voice, iterate with AI, and refine specifications through repeated exchanges. Open offices make that behavior harder. Visual pressure, ambient noise, and the fear of sounding like one is talking to oneself lower both recognition accuracy and concentration. As a result, the design density those people could have produced never fully emerges.

What companies need, therefore, is not a simple binary of “everyone in the office” versus “everyone remote.” They need to acknowledge the value of in-person work while also designing private rooms, quiet booths, telecubes, or flexible home-work arrangements for people who think through speech. Collaborative meetings and high-density individual dialogue with AI do not necessarily thrive in the same physical setting.

This is not merely a matter of employee comfort. It is investment in the means of intellectual production. Just as factories once had to equip skilled workers with the right machinery, AI-era companies must equip people who design through language and inject intent into AI with the right environment. Otherwise, people capable of first-rate results may be misread as merely “a bit difficult to handle.”

Conclusion: AI rewards more than polished summarizers

All of this suggests that in the era of generative AI and AI agents, value does not belong only to people who can write neat short summaries. The ability to bring out unorganized thought, hesitation, edge cases, emotional temperature, and non-negotiable design priorities without losing them may be becoming more important than ever.

Of course, the material does need to be organized eventually. Raw, uncompressed input used as the final instruction creates noise. But there is real value in first getting everything out and then letting AI restructure it. That two-stage workflow can achieve a depth that the older style of “summarize beautifully from the start” often cannot. The key question is not whether the prompt is long or short. It is how much of the design intent survives the transfer.

Viewed this way, the image of the excellent developer changes a little. The person who silently organizes specifications is not the only strong one. The person who thinks by speaking, designs while thinking, and can hand that process to AI is also extremely strong. The tentative label “talking engineer” is simply a way to point at that shift.

And at the far end of that spectrum stands someone like Ichiro Furutachi. If a person with Furutachi-level verbalization could fluently use AI agents and inject multilayered structures and images directly from the mind, the resulting output could exceed many conventional expectations. AI does not only want polished sentences. What it may need most is the raw structure and honest intent inside the human mind before they have been compressed away.