Refreshing Software and Technology Lending Descriptions with Agent-Based AI
We audited and rewrote the descriptions for roughly 300 devices and 150 software titles in the Libraries' technology lending and software catalogs, replacing outdated, vague, or manufacturer-supplied copy with consistent, informative descriptions—aided by an AI agent and a custom-built side-by-side review application.
Overview
The NC State University Libraries has one of the largest and most diverse technology lending collections of any academic library. With hundreds of devices such as cameras, medical models, audio equipment, virtual reality headsets, and more, our website has grown over time to reflect that breadth—but the format and quality of the device descriptions were inconsistent. Many descriptions were functional but vague. Others were composed of language copied from manufacturer websites, which told borrowers little about how a device might actually be used.
The same was true of our 155 software pages. Many entries consisted of a single generic sentence—such as "Use this software for archiving and compressing files"—without naming formats, features, or concrete tasks. Others simply restated the manufacturer tagline.
We set out to bring consistency and usefulness to all 294 device pages and 155 software pages, ensuring that every description follows the same format: two to four sentences about what the device is and what it can do, followed by two to four bullet-pointed example uses. The goal was to help users—students, faculty, and staff—quickly understand whether a device meets their needs. Another benefit of this approach was saving time. By automating the research and first-draft writing, we avoided assigning a large set of repetitive tasks to staff, allowing them to focus on more important parts of their jobs.
We then repeated the same process for the Libraries' 155 software pages, retrieving each entry and having the agent research official documentation to draft descriptions in the same format.
How We Did It
Using a GitHub Copilot agent with Claude Opus 4.6 as the underlying model in VS Code, we retrieved all device pages from the Libraries' web API. For each device, the agent conducted research and produced a draft description following our defined format: a factual explanation of the device's capabilities paired with concrete example use cases. Devices were processed in batches, with descriptions stored in structured JSON files that preserved both the original and the proposed rewrite for comparison.
To evaluate the rewrites, we developed a small web application that displayed each device's original description alongside the AI-generated version ("Suggested Rewrite" in the screenshot below).
Libraries staff read each of the 294 pairs and made one of three decisions:
- approve the rewrite;
- prefer the original; or
- flag the entry for a manual rewrite.
Of the 294 proposed device descriptions, 27 were flagged for additional review or manual rewrite, including six where we preferred the original description. The items flagged for manual rewrite contain requesting or workflow information or need manual attention from a subject matter expert.
For software, staff reviewed all 155 proposed rewrites using the same side-by-side application. Every description was approved for update, with staff making inline edits to refine wording where needed. Four entries were initially flagged for closer review but were ultimately included after revision.
Both the device and software description updates are complete and live on the website.
Examples of revised descriptions
| Item | Original | Rewrite |
|---|---|---|
| Device: Sigma-Aldrich Chem Tutor Modeling Kit | Useful for Organic Chemistry modeling. | This ball-and-stick molecular modeling kit is designed for building 3D models of organic chemistry structures. It can be used to visualize molecular geometry, bond angles, and stereochemistry for coursework or study.
|
| Device: Janome 2212 Sewing Machine | The Janome 2212 is a reliable, easy-to-use mechanical sewing machine - perfect for beginners and those looking for a straight-forward experience. Its light-weight and durable design makes it an ideal machine for home projects, alterations, and crafting. Important Functions: | This mechanical sewing machine has 12 built-in stitch patterns with adjustable stitch length and width. It weighs 10–12 lbs and is portable. It is suited for basic sewing tasks such as hemming, garment construction, alterations, and fabric crafting.
|
| Software: LINDO | LINDO® API creates optimization applications. It allows you to plug the power of the LINDO® solver right into customized applications that you have written. | This mathematical optimization software solves linear, integer, nonlinear, stochastic, and global optimization problems. It is used in operations research, engineering optimization, scheduling, and resource allocation.
|
| Software: Ink/Stitch | (No description) | This free, open-source machine embroidery digitizing platform runs as an extension of Inkscape and converts vector designs into machine embroidery stitch files. It supports multiple stitch types (satin, fill, tartan), includes pre-digitized fonts for lettering, and exports to formats compatible with various embroidery machines.
|
