From Six Years to Six Hours: Prerequisites and Implications of AI-Assisted Development in a Library Context

1 Introduction

iSearch is a locally developed desktop application critical to the workflows of the Metadata Services department at the University of Toronto Libraries (UTL). It supports two core workflows: bibliographic record harvesting, in which the best MARC 21 record is selected from results returned by querying Z39.50 databases; and inventory creation, in which holdings and item-level data is generated for import into UTL's library management system (LMS). iSearch was developed by a librarian who retired without transferring the source code to the University. It cannot be updated, patched, or recovered if it fails. As such, functional requirements and record assessment criteria were documented between 2019 and 2020 in anticipation of a redevelopment project with UTL Information Technology Services (ITS). The project has been deferred by competing institutional priorities, leaving iSearch as an unresolved institutional risk.

In March 2026, the author used Claude to develop several Model Context Protocol (MCP) servers, including one connecting AI agents to Library of Congress linked open data vocabularies[1]. That work built familiarity with AI-assisted development that proved transferable. Emboldened by that experience, the author produced a web-based prototype replacement in approximately six hours. This paper reflects on what made that possible and implications for library practitioners considering AI-assisted development.

2 Prerequisites for AI-Assisted Development

The six hours of AI-assisted development did not happen in a vacuum. Several conditions enabled rapid, directed development.

3 The Development Process

The starting point for redevelopment was the functional requirements and record assessment criteria documented in 2019–2020, supplemented by access to the running iSearch application. A critical piece of guidance came from the retiring developer himself, who pointed to YAZ, Index Data's open-source Z39.50 toolkit[6], as the appropriate library for this work. The idea of using a tournament-based record selection algorithm was similarly informed by the developer.

Rather than building the entire application at once, the author decomposed the development process into discrete tasks, each independently testable. Claude was directed to build a Python/FastAPI backend with a static HTML/CSS/JavaScript frontend. Core components were developed sequentially: Z39.50 querying, the tournament-based record selection algorithm, record assessment against a three-level quality rubric, and the inventory creation workflow. The prototype was tested end-to-end with the exception of SFTP delivery to the production transfer server, which requires network access controlled by ITS.

While Claude was a capable collaborator for translating well-specified requirements into working code, the process was not without friction; human judgment was essential throughout. The output was sometimes subtly wrong in ways only domain expertise could detect: incorrect handling of MARC encoding levels, edge cases in record comparison logic, or assessment criteria applied in the wrong order. Testing was a continuous process, with each component verified against real Z39.50 targets and live bibliographic data.

4 Implications

The iSearch prototype demonstrated that AI-assisted development can produce working, useful software in a compressed timeframe, but building it surfaces implications for institutions seeking to create the conditions for this kind of work to succeed.

5 Conclusion

This case study demonstrates that AI-assisted development can be legitimate and productive for library practitioners under certain conditions. The prototype was enabled by prior documentation, domain expertise, deliberately developed literacies, and proactive pursuit of appropriate tools. None of these prerequisites should be taken for granted; institutions envisioning AI-assisted work must invest deliberately in building them.

The experience also points toward a constructive model for how prototype development can relate to institutional IT processes. Practitioners can frame prototypes as proof-of-concept contributions that accelerate and inform formal redevelopment. As AI-assisted development becomes more accessible, the library community will need to develop shared best practices around coding standards, accessibility, security, and institutional governance; this paper offers one early data point in that conversation.

Acknowledgments

The author gratefully acknowledges the AI Kitchen, co-sponsored by the Office of the Vice-Provost, Digital Strategies and Information Technology Services, University of Toronto, for providing access to Claude Edu, which enabled the development work described in this paper.

References

1. [1] M. Chan, LC Vocabularies MCP Server. 2026. https://github.com/msuicaut/lc-vocabularies-mcp.
2. [2] Library of Congress Network Development and MARC Standards Office, MARC 21 Format for Bibliographic Data. 1999. https://www.loc.gov/marc/bibliographic/.
3. [3] Library of Congress Network Development and MARC Standards Office, MARC 21 Format for Holdings Data. 2000. https://www.loc.gov/marc/holdings/.
4. [4] Program for Cooperative Cataloging, BIBCO Standard Record (BSR) RDA Metadata Application Profile. 2024. https://www.loc.gov/aba/pcc/bibco/documents/PCC-RDA-BSR.pdf.
5. [5] The Carpentries, The Carpentries. 2019. https://carpentries.org/.
6. [6] Index Data, YAZ: Z39.50 Toolkit for C. 1995. https://github.com/indexdata/yaz.
7. [7] G. van Rossum, B. Warsaw, and A. Coghlan, PEP 8 — Style Guide for Python Code. 2001. https://peps.python.org/pep-0008/.
8. [8] A. Kirkpatrick, J. O'Connor, A. Campbell, and M. Cooper, Web Content Accessibility Guidelines (WCAG) 2.1. World Wide Web Consortium (W3C), 2018. https://www.w3.org/TR/WCAG21/.

Appendix A AI Statement

This paper was drafted and revised with Claude (Anthropic, claude-sonnet-4-6), drawing on supplied project documentation and ongoing chat history to provide contextually informed assistance. The author takes full responsibility for the accuracy of all claims.