High-Specification, Not Low-Specification or Minimal-AI: Prompting Practice and Clarity of Instructions to Human Colleagues

Abstract

Among 342 employees across several mid-sized organizations tracked over one fiscal quarter, employees whose logged prompts to a general-purpose AI assistant regularly specified explicit success criteria, constraints, or examples ("high-specification," n=127) had delegated tasks to human colleagues rated clearer by managers and peers (7.6/10) than employees whose logged AI prompts were typically open-ended and under-specified ("low-specification," n=118; 6.2/10, a 23% gap) or employees who used the assistant rarely and continued to delegate to colleagues as before ("minimal-AI," n=97; 6.4/10, a 19% gap). High-specification employees' delegated tasks also drew fewer clarifying follow-up questions per task (1.25) than low-specification employees (1.55, a 24% gap) and minimal-AI employees (1.5), and required rework less often (14% of tasks) than low-specification employees (18%) or minimal-AI employees (17%).

1. Introduction

Advice about delegating work clearly to colleagues is old; what is new is a channel where an employee can practice specifying a task hundreds of times a week and get an immediate, literal readout of whatever ambiguity they left in. A general-purpose AI assistant does not infer unstated intent the way a familiar colleague might, so an under-specified request to it tends to surface as a visibly wrong or generic result rather than a quiet, absorbed guess. It is an open question whether the habit of specifying success criteria that this forces with an AI assistant carries over into how the same person delegates to other people, or stays confined to how they use the assistant itself. We compared instruction clarity to human colleagues across employees who differed in how specifically they had been prompting an AI assistant for work tasks.

2. Method (Summary)

We recruited 342 employees across seven mid-sized organizations (technology, logistics, and professional-services sectors) who used the same general-purpose AI assistant for at least some work tasks during one fiscal quarter. A random sample of each participant's saved work-related prompts (minimum 15 prompts per participant) was coded by two raters blind to the study's hypothesis for whether the prompt specified explicit success criteria, constraints, or a worked example (agreement 89%, disagreements resolved by discussion). Participants were classified as "high-specification" if at least 60% of their sampled prompts met this standard (n=127), "low-specification" if fewer than 30% did (n=118), and "minimal-AI" if they reported using the assistant for work tasks less than once weekly and continued delegating primarily to human colleagues directly (n=97). For each participant, we drew a random sample of tasks they had delegated to a human colleague during the same quarter from the organization's project-tracking tool, and obtained a manager or peer clarity rating (1–10) for each, blind to the participant's group. We also counted clarifying follow-up questions logged per delegated task and whether the task was flagged for rework in the tracker.

3. Findings

High-specification employees' delegated tasks were rated 7.6/10 for clarity on average, versus 6.2/10 for low-specification employees (a 23% gap) and 6.4/10 for minimal-AI employees (a 19% gap); low-specification and minimal-AI employees did not differ meaningfully from each other. High-specification employees' delegated tasks drew 1.25 clarifying follow-up questions per task on average, versus 1.55 for low-specification employees (a 24% gap) and 1.5 for minimal-AI employees (a 20% gap). Tasks delegated by high-specification employees required rework 14% of the time, versus 18% for low-specification employees and 17% for minimal-AI employees.

4. Discussion

The comparison that isolates what matters here is between high- and low-specification AI users, since both groups used the assistant regularly; only the specificity of their prompts differed. That the minimal-AI group's outcomes tracked the low-specification group rather than the high-specification group suggests using an AI assistant frequently is not, by itself, what produced clearer human delegation — the benefit tracked specifically with the practice of stating success criteria and constraints explicitly, a practice the low-specification group had just as much opportunity to adopt with the same tool and did not. Getting an assistant's literal, immediate output back for an under-specified prompt appears to function as repeated, low-stakes practice at noticing one's own unstated assumptions, in a way that a colleague's polite request for clarification rarely forces as directly or as often. The gain here is not simply time saved for the high-specification employees themselves — fewer clarifying rounds and less rework are a cost most directly borne by the colleagues receiving the delegated task, so the benefit accrues to the collaborators as much as to the person who developed the habit.

5. Limitations and Future Directions

Group membership was based on employees' existing prompting habits rather than random assignment, so employees inclined toward precise communication generally may have both specified their AI prompts more and delegated more clearly regardless of any transfer effect. The prompt-coding threshold (60%/30%) was chosen to create clearly separated groups and a different threshold might narrow the observed gap. Clarity ratings came from managers and peers within the same organization as the employee being rated, who were not blind to the employee's general reputation, though they were blind to this study's group classification. A planned follow-up will randomly assign a subset of low-specification employees to a brief prompt-specification training program to test whether the transfer effect holds under assignment.

Editorial Note

This report underwent CfAHC's Rapid Review Protocol, including AI-assisted literature cross-checking and a two-reviewer human sign-off.