chartitec/glaskuser

GlaskUser turns archived user research recordings and transcripts into a set of AI personas you can query like having another interview. The problem it targets is that historical interview data is typically shelved once a project ends, making it hard to answer new questions from the same users or compare responses across a cohort. The core idea is to extract a psychological model from each user's interview material, covering core values, decision frameworks, and inference rules, and ground every AI response to that user's actual corpus. When a question falls outside what was discussed, the persona explicitly says so rather than guessing. This makes it distinct from general AI-generated user profiles, which have no traceability to real people. Getting started requires Claude Code with an Anthropic API key. You run a single setup command (/glaskuser_init) and Claude guides the rest: it installs dependencies, downloads a Whisper speech-to-text model and a semantic search model, transcribes any audio or video, builds a vector index, and extracts psychological models for each user. Supported input formats include audio (.mp3.wav.m4a.aac.ogg.flac.webm), video (.mp4.mov.avi.mkv), transcripts (.txt.pdf.md.docx), and spreadsheets for surveys or usage logs. Audio transcription and semantic search run locally, the indexed corpus is passed to the Anthropic API through Claude Code for questioning. Once set up, /glaskuser_simulate opens a conversation with a single user persona, and /glaskuser_search retrieves raw interview excerpts by topic. Python 3.10 or newer is required. The full README is longer than what was provided.