When people face a hostile or biased evaluator, such as an interviewer who makes a gendered remark, a customer demanding appearance compliance, or a senior colleague pressing a junior one, they need a way to protect themselves. They need to hold a boundary, decline the premise, and keep their composure. Sociologists of dress and presentation call such self-protective signals "social armor." Their evidence suggests the armor works best when the person authors it, because imposed presentation scripts are tiring. AI conversational coaches now let people rehearse these encounters against simulated adversaries. They also introduce a familiar teaching problem: a coach that keeps telling you what to say may produce a trainee who recites its lines instead of building their own.
This note proposes a controlled human study of one specific design choice. An AI coach can name what kind of move to make ("you are accepting their framing; consider redirecting to the task") or supply the words to say. Separately, its help can stay constant or fade over practice. We predict that strategy-only prompts, faded over time, will produce better unassisted boundary-holding in new scenarios. They should also leave the trainee's own voice intact and reduce reported emotional strain. We predict that wording-supplying coaching will look good during practice and transfer worse, behaving like a conversational version of a corporate "look book."
Core Claim
The bridge between the two literatures is a structural analogy that has not been tested. On the identity side, Kim's survey of cabin crew found that autonomy over grooming standards predicts job vitality, dedication, and commitment, and through them service performance. Kim treats mandated look books as a source of emotional labor. Shan's synthesis of gender theory describes apparel as defensive armor during identity vulnerability, but notes the field rests on theory and retrospective self-report rather than moment-by-moment measurement.
On the teaching side, Abdelshiheed et al. showed in a logic tutor that timed nudges prompting a strategy switch work better than presenting the strategy. The nudges leave the decision with the learner. Six weeks later in a different tutor, nudged students scored 91.9 on transfer versus 70.3 for controls, and they more often chose to solve problems unaided. Wilhelm et al.'s CommCoach study reports that managers valued in-the-moment AI feedback. The same managers worried that always-on coaching pushed them toward following an algorithmic script rather than developing their own style.
Our claim is that "nudge versus present" from tutoring corresponds to "self-authored versus mandated presentation" from the armor literature. We further claim that in identity-salient conversation this distinction decides whether coaching produces transferable self-protection or dependent script compliance. What is new here is a randomized test that separates the content of coaching (strategy name versus wording) from its schedule (constant versus fading). The test measures unassisted transfer, how often trainees echo the coach's phrasing, and emotional strain. None of the cited work measures these together.
Research Program
We will build a practice environment with two language-model agents, following CommCoach's split between an interlocutor agent and a coach agent. The interlocutor plays scripted adversarial personas. Each persona applies identity-salient pressure through demeaning framing, appearance or name-related remarks, and premise-loading questions, while staying within ethics-board limits. The coach watches each learner turn. It classifies whether the learner is holding a boundary, redirecting, or accommodating the hostile premise. It then fires a prompt in one of two styles.
A strategy-only prompt names a move from a fixed taxonomy and gives no example wording. The taxonomy has five moves: name the behavior, decline the premise, redirect to the task, state a limit, and exit.
A wording-supplied prompt gives a ready-to-use sentence for that same move.
The two prompt styles share the same trigger classifier, so they differ only in content.
Fading follows the decaying-threshold form from SWITCH. SWITCH is a distillation method in which a teacher model takes over generation when its disagreement with the student exceeds a threshold that shrinks over training. In SWITCH the teacher ends up writing about 53% of tokens. Here we borrow the schedule, $\tau_t = \tau_0 e^{-\lambda t}$, where $\tau_t$ is the confidence the coach's accommodation detector must exceed before it intervenes at practice step $t$. $\tau_0$ sets the starting sensitivity and $\lambda$ the rate of withdrawal. Help therefore becomes rarer as practice proceeds.
SWITCH is a constraint as well as a source. It shows that takeover helps a model student avoid compounding errors. Our hypothesis is that, for human identity work, takeover (the wording-supplied arm) carries a cost that SWITCH's setting cannot reveal.
The learner-state concern comes from Gao et al.. They found that a "best-overall" tutoring policy produced negative normalized learning gains (−12 and −20) for the weakest subgroups, while subgroup-selected policies produced +24 and +32. We do not build their policy selector. We instead preregister a moderation analysis using two pretest measures, prior experience of identity-based workplace hostility and baseline boundary-holding skill. This tests whether the best arm differs for more vulnerable trainees.
Baselines
The naive reference point is practice-only. Participants run the same number of adversarial scenarios with the same interlocutor and no coach. This tests whether rehearsal alone does the work.
The rigid-guidance baseline gives a static script card of model responses before practice. It is the conversational look book, the condition Kim and Shan predict will cost emotional effort.
The post-hoc baseline gives a written summary after each scenario, covering the same moves the coach would have flagged. It is the obvious way to protect autonomy during the conversation.
The strongest existing approach is CommCoach-style always-on, wording-supplied in-situ coaching. It is the closest deployed system and the one practitioners preferred. For this baseline, no reported result was found on Lacuna for CommCoach or any conversational boundary-training task; Wilhelm et al. report only qualitative themes from 17 managers. The only quantitative anchor for nudging comes from a different domain: Abdelshiheed et al.'s transfer of 91.9 for nudges versus 70.3 for controls and 89.3 for the strong-strategy group. It shows the effect size nudging can reach. It is not a target number for this task.
The key ablation is the 2×2 of strategy-only versus wording-supplied crossed with always-on versus faded. Comparing the two faded arms isolates prompt content with schedule fixed. Comparing the two strategy-only arms isolates fading.
All seven arms use the same base model, the same six practice personas, and the same 40-minute practice budget. Each arm's coach prompt gets the same number of pilot tuning rounds (we plan five) against a shared pilot pool. Participants are randomized after the pretest.
The proposal beats the strongest baseline if faded strategy-only coaching exceeds always-on wording-supplied coaching on the primary outcome, unassisted transfer boundary-holding, by a preregistered margin of 0.3 SD. It must do so with lower script-echo, and without being worse than practice-only on emotional strain.
Why not just…?
Why not just let people practice against the adversarial persona with no coach? The tutoring and programming evidence says unguided novices flounder. Bender et al. found open-ended Scratch work produced trial-and-error "bricolage" and higher cognitive load (3.90 versus 3.12 under Parsons puzzles). MindScratch found only 54% of baseline students met objectives. Social rehearsal may differ, though, since people already have some conversational repertoire. The practice-only arm settles it. If it matches faded strategy-only coaching on transfer, coaching is unnecessary and we will say so.
Why not just give feedback afterward? Post-hoc feedback fully protects in-the-moment autonomy. CommCoach users preferred in-situ feedback, but preference is not learning. SWITCH's argument that errors compound over a sequence comes from model generation, not human conversation. This is a real threat. If the post-hoc arm matches faded strategy-only on transfer and voice retention, the simpler design wins.
Why not just use a fixed fading schedule with wording supplied, since fading alone might prevent dependency? This is the most plausible competitor. It is the reason the study crosses content with schedule instead of testing a single bundled system. If faded wording-supplied coaching equals faded strategy-only coaching on transfer and script-echo, our central claim about authorship is wrong. The finding would reduce to "fade your help," which is already standard teaching advice.
Why not just hand people a good script? Kim's look-book findings, Shan's "suit trap," and Abdelshiheed et al.'s finding that presented strategies underperform nudges all predict a cost. None of them measured it in conversation. The script-card arm measures it directly, through both transfer and emotional strain.
Proposed Experiments
The main study recruits about 280 early-career professionals, including service-sector workers as a nearer analog to Kim's cabin crew, and assigns 40 to each arm. It has three parts. A pretest scenario with a novel adversarial persona comes first. Two practice sessions a few days apart follow. Unassisted transfer is tested in two scenarios: one immediately, with a new persona and a new pressure type, and one about four weeks later. The delayed test mirrors Abdelshiheed et al.'s delayed transfer design.
The primary outcome is boundary-holding on transfer. Two blinded human raters score it per turn with the five-move taxonomy plus an "accommodated the hostile premise" code, and we report inter-rater agreement. An automatic judge is used only if it matches the raters on a held-out set.
Script-echo is the share of the learner's transfer turns that closely match coach-supplied or script-card phrasings, measured by n-gram overlap. A complementary voice-retention score is the stylistic similarity of transfer turns to the learner's own pretest turns. Emotional strain and self-efficacy are self-reported after each session. We also record a behavioral confidence measure modeled on Abdelshiheed et al.: in a final optional round, does the participant choose to practice with or without the coach?
Three decisive checks follow from the design. If wording-supplied coaching shows lower script-echo at delayed transfer than at practice, while strategy-only coaching shows no such gap, the dependency mechanism is supported. If the two faded arms do not differ on transfer, the authorship hypothesis fails. If the moderation analysis shows the most vulnerable trainees benefit more from wording-supplied help early, that would partly vindicate Gao et al.'s heterogeneity concern. It would point to a follow-up that adapts prompt content by learner profile rather than only fading it.
Ethics deserves explicit care. Personas avoid slurs. Participants can end any scenario. A debrief follows each session.
References
- A Study on the Effects of Autonomy of Cabin Crew’s Image Making Standards on Job Enthusiasm, Service Performance and Organizational Performance
- An Examination of the Effects of Abandoning the Symbolic Definition of Gender Through Apparel in Accordance with Contemporary Gender Theories
- The Power of Nudging: Exploring Three Interventions for Metacognitive Skills Instruction across Intelligent Tutoring Systems
- How Managers Perceive AI-Assisted Conversational Training for Workplace Communication
- SWITCH: Studying with Teacher for Knowledge Distillation of Large Language Models
- Get a Head Start: On-Demand Pedagogical Policy Selection in Intelligent Tutoring
- Integrating Parsons puzzles within Scratch enables efficient computational thinking learning
- MindScratch: A Visual Programming Support Tool for Classroom Learning Based on Multimodal Generative AI