Research Sketches

Workflow experiments, analysis techniques, and methodological notes: ways AI has proven useful, uneven, or surprisingly limited in actual scholarship.

How to read these

Treat each research sketch as a field note from an experiment. The point is not that a tool solved the problem, but what it made possible, what it got wrong, and what expertise was still needed around archives, metadata, data ethics, accessibility, privacy, method, or research computing.

Date Apr 2026 Status tested Type processing sources Time downloaded document images; two hours to set up; automated run of ~12hr/register Level researcher Topics archivesbig datapaleography
Pipelines for Medieval Handwriting Recognition

Pipelines for Medieval Handwriting Recognition

ExperimentTo create an AI agent to work with Gemini and Claude to bulk process 300 images of archival documents and enable full-text search of medieval handwriting.

Results
  • built an agentic pipeline for bulk document processing
  • combined multiple LLMs to improve transcription accuracy
  • enabled full-text search of handwritten archival sources
Date Sep 2026 Status tested Type LLM orientation Time less than 10 minutes Level any Topics promptingmodel tierscitations
Initiate Research with an Annotated Bibliography

Initiate Research with an Annotated Bibliography

ExperimentThis sketch demonstrates how to construct a prompt for a high quality annotated bibliography.

Results
  • a sophisticated and reliable annotated bibliography to initiate research
Date Apr 2026 Status tested Type data work Time less than 1 hour Level any Topics AI-assisted codingmapsagentic AI
Photos to Map Pins

Photos to Map Pins

ExperimentCreate an interactive map with pins for hundreds of photos, using GPS metadata already embedded in your phone's images — in under an hour.

Results
  • extracted GPS metadata from image files
  • built a map visualization with AI-assisted coding
  • presented geolocated data in a public-facing format
Date Sep 2026 Status tested Type LLM orientation Time less than 10 minutes Level any Topics promptingmodel tiershallucinations
Effect of Model Tiers on LLM Responses

Effect of Model Tiers on LLM Responses

ExperimentThis sketch shows the difference in response quality baseed on the LLM model tier.

Results
  • hallucinations occur far less frequently with higher tier models
Date Aug 2026 Status tested Type citation work Time less than 10 minutes Level any Topics citationspromptingagentic AI
Table of Contents to Zotero

Table of Contents to Zotero

ExperimentWith one prompt get entries in Zotero from a book of essay chapters.

Results
  • extracted data from the Table of Contents
  • built a file in BibTeX format
  • imported to Zotero
Date Apr 2026 Status tested Type data work Time 30–60 min Level any Topics 3D printingmaterial culture
Generate 3D Prints from 2D Drawings

Generate 3D Prints from 2D Drawings

ExperimentAI can transform a historical line drawing into a 3D-printable file, adding a tactile dimension to research that images alone can't provide.

Results
  • generated 3D-printable files from 2D historical images
  • reconstructed material culture objects for research
  • incorporated tactile elements into research presentations
Date Aug 2026 Status tested Type writing feedback Time one prompt per reviewer; an afternoon to work through the responses Level researcher Topics writingpeer reviewprompting
A Panel of Specialist Readers

A Panel of Specialist Readers

ExperimentAsk AI to review a draft article separately as each of the specific readers it will actually face — subfield expert, adjacent specialist, volume editor, fellow contributor.

Results
  • four distinct reviews of the same draft, each from a named vantage point
  • insights from every perspective that I had not considered
  • substantive revisions to the article

Browse all sketchbook tags →