The Instruction Gap: Why the Real AI Skill Nobody Is Teaching Is Following Directions
- David Borish
- 17 minutes ago
- 5 min read
Following Instructions Is the New Moat
A product manager at a mid-sized SaaS company recently described spending three hours with Claude Code trying to build an internal data pipeline. The AI did not fail. It produced a complete, working implementation along with a precise sequence of steps to deploy it: configure the environment variables, authenticate with the cloud provider, install the dependencies in a specific order, run two terminal commands. She got through the first two steps, got confused by the credential format on the third, closed the terminal, and opened a ticket for engineering. The pipeline sat unbuilt for two weeks, not because the AI failed, but because the human receiving its output could not complete the steps it handed back.
The Thesis
There is a version of the current AI moment that focuses almost entirely on prompting, on how to phrase requests, how to give context, how to coax better outputs from a model. That framing captures something real. Getting good output from an AI agent requires skill and discipline. Addy Osmani, an engineering lead at Anthropic, wrote recently that effective use of these tools is "difficult and unintuitive" and requires learning new patterns even for experienced developers. Approximately 90% of Claude Code's own codebase is now written by Claude Code itself, and the people using it most effectively are not just better prompters. They have developed structured workflows for what happens after the AI responds.
But prompting is only half the loop. Once the AI returns a response, something else has to happen: a human being has to execute the steps. That is where most of the failure occurs, and it is a failure mode that the prompting narrative has largely ignored. Prompting is what you give the AI. What matters just as much is what the AI gives back, and whether the person receiving it can carry it through to completion.
What the Research Shows
Cognitive scientists have spent decades studying why people fail to complete multi-step instructions, and the findings are consistent. Working memory is the primary constraint. A 2020 review in the American Journal of Pharmaceutical Education found that working memory capacity is the central variable in instruction-following ability, and that people with lower working memory are particularly prone to starting sequences and failing to complete them. This is not about intelligence or motivation. It is about the cognitive load imposed by holding multiple steps in mind while executing each one.
Susan Gathercole and Tracy Alloway documented a related pattern in their research on working memory and learning: both children and adults with limited working memory tend to begin multi-step tasks and then lose track before finishing. The breakdown usually happens somewhere in the middle, not at the start. The first step is manageable. The fifth step, after holding steps two through four in mind while executing them, is where people drift.
The modality of instruction matters too. A 2015 study in Scientific Reports by Yang, Allen, and colleagues found that demonstrated instructions, where someone physically shows you each step, are retained significantly better than written or spoken instructions alone. Reading a sequence of steps in a chat window is close to the worst possible format for retention. It is linear, it is passive, and once you close the window to open your terminal, the instructions are gone.
Jaroslawska, Gathercole, Allen, and Holmes found in a 2016 study in Memory & Cognition that physically performing each step at the moment of instruction dramatically improved both retention and completion. The implication for AI-assisted workflows is practical: people who execute each step immediately as they read it perform better than people who read the full list first and then try to execute. The habit of reading ahead undermines the very completion it is meant to prepare for.
Stanley Milgram, in his well-known authority research, identified something else worth noting: instruction-following is situational, not dispositional. People are more likely to complete instructions when an authority figure is present to enforce completion. Remove the authority and the completion rate drops. The AI has no authority. It issues instructions and waits. There is no follow-up, no checkpoint, no one standing in the room. For many people, that absence is enough to stall them.
The Prompting Narrative Is Incomplete
The popular story about AI productivity goes roughly like this: the people who will benefit most are the ones who learn to prompt well. Write clearer requests, give better context, iterate on outputs, and the AI becomes a force multiplier. That story is not wrong, exactly, but it stops too early in the workflow.
Fortune's coverage of the emerging "supervisor class" describes how developers are shifting from writing code to reviewing and directing agent outputs. The piece frames this as an elevation of the role: less syntax memorization, more judgment. What it does not fully address is the step between direction and completion. Reviewing agent output is one skill. Following the operational instructions that output contains, particularly when those instructions involve unfamiliar tools, systems, or credential chains, is a different skill entirely, and it is one that organizations have not trained for.
The teach-back method, a standard technique in health literacy and patient education, offers a useful frame. It is not enough to give someone instructions; you ask them to repeat the instructions back in their own words before they leave the room. The reason this technique was developed is that comprehension and execution are not the same thing. People nod at instructions they cannot complete. The AI does not know to ask.
The Enterprise Implication
Most enterprise AI rollout failures are discussed in terms of model selection, data readiness, or change management. Those factors are real. But a significant portion of the friction happens at a more granular level: individual contributors receiving AI-generated workflows and not completing them.
This is measurable in lost time, in re-routed tickets, in tasks that land back in engineering queues after a non-technical user stalled on step three of a deployment sequence. Organizations tracking AI adoption by seat count or by the number of prompts submitted are measuring the wrong thing. The metric that matters is completion: what percentage of AI-generated instruction sequences actually get executed end to end by the person who received them?
The answer, in most organizations, is lower than assumed. And the gap between "AI generated a plan" and "the plan was implemented" is where the return on AI investment disappears.
Who Survives
The workforce displacement conversation tends to focus on which roles AI can perform. The more proximate question, right now, is which people can work effectively alongside AI outputs, not by generating better prompts, but by completing the operational sequences the AI returns.
The profile is not the best programmer. It is not even necessarily the most technically literate person in the room. It is the person who can read a ten-step deployment sequence, hold it in working memory, execute each step sequentially without losing track, recognize when a step has produced an unexpected result, and adapt without abandoning the task. These are the same attributes that predict success in surgical training, in manufacturing environments with complex procedures, and in any domain where multi-step protocols must be followed under conditions that produce cognitive load.
That skill has never been formally valued in knowledge work because knowledge work did not previously require it at this level of specificity. A content strategist did not need to configure API keys. A marketing analyst did not need to install dependencies in a specific order. They do now, if they want to use the tools that have become central to competitive output.
The organizations that figure this out first will stop asking "who can prompt well" and start asking "who can follow through." Those are related questions, but they are not the same question. And the second one is, increasingly, the one that determines whether an AI investment produces anything at all.
Comments