Conversation model
Assistant has its own model. Your Coder projects keep their choices.
Her voice
Estelle is your Black British assistant from London. Audition these local British English voices to choose her sound.
Learning your rhythm
Only submitted voice turns contribute. Recordings aren’t retained.
Suggested learning
Check what Estelle understood and when it applies before saving.
Teach Estelle
Saved preferences guide future replies. This doesn’t retrain model weights. Older chat stays in history; Estelle uses recent context and can look up earlier messages in the same chat.
Saved learning
Coding improvement
Estelle scores recorded checks and isolated worker candidates. Models may suggest reusable lessons, but only accepted guidance reaches future workers. Raw prompts, source files and check logs aren’t copied into lessons.
Historical coding baseline
Use prior Codex session metadata, registered project structure and reviewed Project Manager results to recognize familiar work. Transcripts, prompts, commands, source contents and credentials are excluded. Historical context never becomes accepted guidance automatically.
Completed result learning data
These bounded task and project reports include outcome, process, verification and source provenance. Coder can retrieve them as context; they do not retrain model weights or become accepted guidance.
Baseline projects and recent sessions
Provider evidence
Repeatable evaluations
Define one assignment and its acceptance criteria, then use the same case across providers. Attach only the outcome that actually ran.
Coding lessons to review
Accepted coding guidance
On your iPhone
Open Estelle in Safari, tap Share, then Add to Home Screen. Your server needs to stay on and connected.