Huge thanks to my wonderful collaborators Kathryn Yurechko, Pranati Dani, @cqz.bsky.social and @axz.bsky.social.

Full details here ➡️ arxiv.org/pdf/2409.03247

All three strategies struggled with iterative refinement.

Interestingly, participants adopted hybrid approaches when iterating on their prompt filters – like providing examples as in-context examples or writing rule-like prompts.

Despite 🤖LLM prompting’s better performance, participants preferred mixed strategies to create their filters.

For example, when their preferences were ill-defined but intuitive, 🔎labeling examples was considered the easiest way. (🧵4/N)

To answer this question, our study had 37 non-programmers create personal content filters using these three strategies. (🧵3/N)

Existing content filter tools often expect lay people to work in a single, long setup session. Yet users engage with social media in short, everyday sessions.

How can we support social media users to more easily create and iterate on their filters? (🧵2/N)

Can LLM prompting help social media users create and iterate on their content filters more easily?

In our #CHI2025 paper, we compared in an experiment three authoring strategies:
🤖 Prompting LLM
🔎 Labeling examples for ML classifiers
📐 Authoring keyword rules

(🧵1/N)