GUARD gets the Gist!
When a large language model (LLM) characterizes an institution, it often substitutes the category's gist for the institution's own verbatim record, the very thing a funder, partner, or parent carries into a decision about a university, and in this instance, Morgan State University (MSU). For MSU, that substitution is not a surface error but a reproduced stereotype. This is hard to catch because no single sentence is false; it is just not fully inclusive of the amazing facts about our institution! The GUARD.md skill (Ground, Unbundle, Anchor, Resolve, Distill) was designed to be uploaded to Claude, Anthropic's LLM, and requires the LLM to verify entity-specific evidence and state an institution's own gist before any category label is applied.
GUARD.md A free Claude skill file for correcting categorical gist substitution, a bias pattern in how large language models characterize institutions“Morgan State University is a public historically Black research university… the largest HBCU in Maryland…”
“Morgan State University holds the R2 designation… Maryland's Preeminent Public Urban Research University… $104.4M in FY25 sponsored research.”
Background
Correct facts, but there was more to the story!
During an internal review of the TRAILS Bi-Annual report, generated by the LLM Claude, from Anthropic, a committee found that when Claude was prompted to describe Morgan State's role in a research collaboration carrying more than $20 million in grant funding, the LLM output characterized the university as an access-oriented recipient of support rather than as a well-resourced contributor. Wang et al. (2024) documented this same pattern, group-level stereotype substituting for case-specific evidence, as systemic across large language models.
The characterization omitted Morgan's Carnegie R2 classification, its designation under Maryland law as the state's Preeminent Public Urban Research University, its $104.4 million in FY25 sponsored research commitments, and its path to R1 status under the Transformation Morgan 2030 plan. It also omitted named research centers, including the Center for Equitable AI and Machine Learning Systems (CEAMLS), the Center for Research and Education in Microelectronics, and the Cybersecurity Assurance and Policy Center. Morgan is one of only 11 HBCUs nationally to hold high research activity classification.
The pattern surfaced clearly in a section of the broadening participation report, and the committee flagged it as a page of concern. During the meeting, Valerie Riggs, Associate Professor at Morgan State University, shared that she had seen this same bias firsthand in LLM output when she worked with faculty and students at a rural university in Central Africa during her Fulbright Specialist project on AI capacity building in Equatorial Guinea. Colleagues were intrigued, and as a result, the NSF TRAILS Broadening Participation Committee appointed her as principal investigator for a research response to the bias identified in the Bi-Annual report. Dr. Riggs recruited an MSU team of scholars to include project lead and developer, computer engineering doctoral student Emmanuel Masa-Ibi; red team leader Daniel Onyejiekwe; and red team undergraduates Ato Bhatta and Yosmehily Segura-Silverio, who together developed and tested the GUARD.md skill.
GUARD was designed to interrupt that substitution.
Watch
The GUARD explainer video
Overview
Ground, Unbundle, Anchor, Resolve, Distill
The GUARD.md skill, developed for use with Claude, Anthropic's LLM, requires the LLM to verify entity-specific evidence before a category label is applied. On a 17-prompt Tier 1 benchmark of HBCU characterization tasks, GUARD-Controlled raised mean gist displacement from 0.79 to 2.39 (Cohen's d = 1.70, p < .0001), reduced failure-signature prevalence from 51.0% to 0.0% (χ² = 31.5, p < .0001), and increased explicit bottom-line articulation from 18.4% to 54.9% (χ² = 12.8, p < .001). A generic control condition produced no significant improvement, isolating the effect to the skill itself.
Each prompt runs under four conditions: an unmodified baseline, GUARD applied inline, a GUARD-Controlled variant that validates its own output against a structural checker before it's accepted, and a generic control. Each response is graded blind on move sequencing, verbatim density, verbatim displacement, and gist displacement, with support from TRAILS.
How it works
Facts first, category last
The GUARD.md skill gives the LLM five moves to force a specific check at each step, so the entity's own record sets the terms of the answer instead of borrowing them from the category it belongs to, which can be stereotypical and contribute to bias in the outcome. Here are the five moves of G/U/A/R/D below:
Open with the specific entity, not its category. No topic sentence about HBCUs in general before Morgan State has been named.
Name the assumption the category would otherwise supply automatically, so it can be checked instead of quietly used.
Pull in real, dated, sourceable facts about the entity itself: classifications, figures, named programs, current leadership.
Where the category assumption and the entity's facts disagree, the facts win, and the description is rebuilt around them.
Close with a plain-English bottom line the reader can actually carry away, not just a list of facts.
GUARD's framework combines fuzzy-trace theory, which explains why models default to gist, with prompt-based correction methods like chain-of-thought and Constitutional AI, which supply the mechanism for redirecting them.
Download
Compare the outputs, then download the skill
These are unedited model responses from the GUARD evaluation set: the same prompt, run once without GUARD and once with it. Pick an institution to compare.
Run GUARD on your own prompts
GUARD is a plain-text instruction set, not an app to install. It works in Claude, ChatGPT, Gemini, and most other chatbots. Download a skill file below and follow the setup guide for your platform. Once it's loaded, GUARD applies automatically whenever you ask the model to describe, characterize, or profile a specific entity. No extra prompting required.
A single-pass filter. GUARD writes its answer straight through, applying the five moves as it goes.
A human-in-the-loop version. GUARD-Controlled loops back to you three times, checking its output against GUARD's criteria to make sure it's detailed enough for what you needed.
The skill files are plain Markdown, readable before you install them and safe to check into your own project. The setup guides walk through Claude, ChatGPT, and Gemini for GUARD-inline, and API/Python setup for GUARD-Controlled.
Sample use cases for GUARD
Consider a partnership brief in which Claude is asked to describe a rural community college serving a Hispanic-Serving Institution (HSI) population. Without GUARD, the response may default to the MSI category itself, emphasizing access and underserved communities while omitting what the college is actually doing, such as a bilingual AI workforce training program built for its region or a growing dual-enrollment pipeline. With GUARD, Claude verifies the record before composing the summary, so the brief reflects the institution itself rather than the category it belongs to.
Consider a journalist on deadline, using Claude to draft background before interviewing a small women's college for a feature. Without GUARD, Claude's response can default to the category, characterizing the institution as close-knit and historically women-focused while leaving out the research program or graduate outcomes that would have shaped sharper interview questions. The summary reads as complete, so it becomes the frame for the piece, and a misrepresented institution becomes a missed angle. With GUARD, Claude verifies the institution's own record first, giving the journalist an accurate starting point for the story.
Load GUARD into Claude
- Download the .md file above.
- Open Claude and go to Settings → Capabilities → Skills (or your local Skills folder if using Claude Code/desktop).
- Upload the file, or drop it into your local Skills folder.
- Start a new chat. GUARD applies automatically whenever you ask Claude to describe or characterize an entity.
GUARD is a research project funded through NSF TRAILS (NSF Award No. 2229885), the Institute for Trustworthy AI in Law and Society, and led by faculty and student researchers at Morgan State University. This page accompanies the research paper describing the GUARD skill and its empirical evaluation.
Abstract
GUARD: A Prompt-Level Discipline for Correcting Categorical Gist Substitution in Large Language Model Characterizations of Institutions
Large language models frequently characterize institutions by activating a familiar category prior and describing entities primarily through that lens, even when the category framing understates or contradicts the entity's specific verifiable record. We call this failure mode categorical gist substitution: the model retrieves the categorical summary in place of the entity-anchored one. We introduce GUARD (Ground, Unbundle, Anchor, Resolve, Distill), a prompt-level intervention grounded in fuzzy-trace theory that requires the model to perform five ordered moves before generating its final characterization. We evaluate GUARD on Tier 1 of a curated benchmark of 17 institutional-characterization prompts targeting historically Black colleges and universities, using Claude Fable 5 as the target model. Three graders independently scored 203 responses across four configurations: baseline, control, GUARD-inline, and GUARD-Controlled. GUARD-Controlled, which pairs the discipline with a deterministic checker that validates structural adherence before the model composes its final narrative, produces the largest and most consistent effects, raising mean gist displacement over baseline by a factor of 2.03 (0.79 to 2.39 on a 0 to 3 scale; Cohen's d = 1.70, p < .0001), reducing failure-signature prevalence from 51.0% to 0.0% (χ² = 31.5, p < .0001), and increasing explicit bottom-line articulation from 18.4% to 54.9% (χ² = 12.8, p < .001). A generic "be specific" control condition produces no significant improvement over baseline, indicating that the effect is attributable to the discipline itself rather than to prompt length or instructional emphasis. We report inter-rater reliability, name the study's methodological limitations honestly, including a small grader panel, single-tier scope, and single-model target, and situate GUARD within the broader landscape of prompt-level and deterministic-validation approaches to correcting language model behavior.
Keywords categorical gist substitution · fuzzy-trace theory · prompt-level intervention · GUARD-Controlled · Tier 1 HBCU benchmark · inter-rater reliability
Meet the team
The people behind GUARD
GUARD is built and evaluated by a small team supported by TRAILS.
