The Prompt That Actually Works vs. the One That Doesn't
"Write a presentation for small business owners about our new service." Feed that into any AI tool and you'll get five generic slides - vague enough to apply to any product, useful to no one in particular. Now try: "Outline five slides introducing our new service to small-business owners. Use plain English. End with a next step." Same task, same tool, dramatically different result. The gap between those two prompts is what Direction actually measures - and it's larger, and more consequential, than most people assume.
Why specificity isn't a nice-to-have
A landmark 2022 study by Google Research, published at NeurIPS conference, tested what happens when a prompt walks a model through reasoning step by step instead of just asking for an answer. On a benchmark of grade-school math problems, a large model's accuracy jumped from around 18 percent with a plain prompt to over 57 percent with a structured, worked-example prompt - more than triple, from a wording change alone, no change to the underlying model. The finding reshaped how the field thinks about prompting: the model's raw capability didn't change, but the accuracy anyone could actually extract from it did, dramatically, based purely on how the question was framed.
The gap this closes - and who it closes for
A National Bureau of Economic Research study of nearly 5,000 customer support agents using an AI assistant found that AI use raised issue resolution by 14 percent on average — but that number hides a much sharper story. Newer, less experienced agents saw gains as high as 35 percent, while the most experienced agents barely improved at all. The AI assistant was, in effect, supplying the structure and framing that veteran agents already carried in their heads. Direction is exactly that structure - and the people with the least of it built up through experience are the ones who benefit most from having it supplied well.
India has the usage, not yet the training
This gap matters more in India than almost anywhere else, for a simple reason: usage is already near-universal, but formal training in how to use it well is not. Microsoft and LinkedIn's 2024 Work Trend Index found 92 percent of Indian knowledge workers already use AI at work, well above the 75 percent global figure. But a 2025 ANSR survey of over 3,000 professionals across India's Global Capability Centers found that more than 70 percent of them learned AI skills independently — YouTube, open courses, trial and error - while only about a third had access to any formal employer training. India has, in effect, the highest-usage, least-structured AI workforce of any major market: enormous daily practice, with wide, untested variance in whether that practice actually includes good Direction.
What separates the two prompts, concretely
Go back to the two prompts that opened this piece. The second one wins for four specific, repeatable reasons: it names exactly what to produce ("five slides," not "a presentation"), it states the audience ("small-business owners"), it sets the tone and format ("plain English"), and it gives a concrete constraint that shapes the ending ("end with a next step") instead of leaving the model to guess what "good" looks like. None of this requires technical skill. It requires knowing that a vague ask produces a vague, average-case answer, and a specific one narrows the model down to the specific case you actually needed.
The takeaway
Direction is learnable in an afternoon, in a way Verification and Discernment often aren't - there's no ambiguity to sit with, just a habit to build: before you send a prompt, check whether it names the output, the audience, the format, and the one constraint that actually matters. If it doesn't, you're not testing the model's capability. You're testing how well it can guess what you meant - and the data above says that's a bet worth stop making.
Why specificity isn't a nice-to-have
A landmark 2022 study by Google Research, published at NeurIPS conference, tested what happens when a prompt walks a model through reasoning step by step instead of just asking for an answer. On a benchmark of grade-school math problems, a large model's accuracy jumped from around 18 percent with a plain prompt to over 57 percent with a structured, worked-example prompt - more than triple, from a wording change alone, no change to the underlying model. The finding reshaped how the field thinks about prompting: the model's raw capability didn't change, but the accuracy anyone could actually extract from it did, dramatically, based purely on how the question was framed.
The gap this closes - and who it closes for
A National Bureau of Economic Research study of nearly 5,000 customer support agents using an AI assistant found that AI use raised issue resolution by 14 percent on average — but that number hides a much sharper story. Newer, less experienced agents saw gains as high as 35 percent, while the most experienced agents barely improved at all. The AI assistant was, in effect, supplying the structure and framing that veteran agents already carried in their heads. Direction is exactly that structure - and the people with the least of it built up through experience are the ones who benefit most from having it supplied well.
India has the usage, not yet the training
This gap matters more in India than almost anywhere else, for a simple reason: usage is already near-universal, but formal training in how to use it well is not. Microsoft and LinkedIn's 2024 Work Trend Index found 92 percent of Indian knowledge workers already use AI at work, well above the 75 percent global figure. But a 2025 ANSR survey of over 3,000 professionals across India's Global Capability Centers found that more than 70 percent of them learned AI skills independently — YouTube, open courses, trial and error - while only about a third had access to any formal employer training. India has, in effect, the highest-usage, least-structured AI workforce of any major market: enormous daily practice, with wide, untested variance in whether that practice actually includes good Direction.
What separates the two prompts, concretely
Go back to the two prompts that opened this piece. The second one wins for four specific, repeatable reasons: it names exactly what to produce ("five slides," not "a presentation"), it states the audience ("small-business owners"), it sets the tone and format ("plain English"), and it gives a concrete constraint that shapes the ending ("end with a next step") instead of leaving the model to guess what "good" looks like. None of this requires technical skill. It requires knowing that a vague ask produces a vague, average-case answer, and a specific one narrows the model down to the specific case you actually needed.
The takeaway
Direction is learnable in an afternoon, in a way Verification and Discernment often aren't - there's no ambiguity to sit with, just a habit to build: before you send a prompt, check whether it names the output, the audience, the format, and the one constraint that actually matters. If it doesn't, you're not testing the model's capability. You're testing how well it can guess what you meant - and the data above says that's a bet worth stop making.
Next Story