We implemented Bonsai on Bluesky and conducted a two-phase, multi-week study with 15 participants. This deployment allowed us to observe how people used intentional feedbuilding in practice, and how it compared to their experiences with engagement-driven defaults.

In the ranking stage, Bonsai orders the curated content using criteria derived from the user’s stated intent—rather than predicted engagement—making the logic behind feed prioritization transparent and directly aligned with user goals.

In the curating stage, users can apply natural language prompts (e.g., “focus on recent policy updates” or “exclude promotional posts”) to filter and organize the sourced content, ensuring the feed reflects users' goals / preferences. Each prompt is fed into an LLM that individually ranks content.

In the sourcing stage, Bonsai gathers a wide pool of candidate posts aligned with user goals. Users can refine this stage by editing sources (adding or removing accounts, hashtags, or feeds) to shape where their feed draws content from.

In the planning stage, users express their goals in natural language (e.g., “updates on AI policy” or “posts from close colleagues”). Bonsai translates these goals into structured representations that guide the subsequent sourcing, curating, and ranking of content by providing initial suggestions.

With Bonsai, users can articulate what they want from their feeds (e.g., tracking research, staying informed on a policy area, or connecting with a community) and the system procedurally builds a feed that reflects those intentions in four steps, which we discuss below.

Bonsai sits within a broader debate on recommender systems. While TikTok or Meta optimize (mostly) for attention capture, Bonsai explores what feeds look like when personalized for user intent. Under our taxonomy, it explores the design space of "intentional" and "personalized" feeds!

Fun to see an OLPC XO in the wild!