How a check-in is built

1. You write freely

   One text box. No forms, no mood emoji grid, no streaks to keep you hostage.

2. It names the feeling

   A language model reads the passage and proposes one feeling label with an intensity, plus the alternative it considered and rejected.

3. It quotes you

   The reflection carries your own sentence, unchanged. If it cannot find one, it says so instead of guessing.

4. One small step

   A single, doable action for the next hour: four minutes of breathing, a two-line journal prompt, a walk, a message to a person.

5. Crisis words are handled by rules

   A fixed word list runs before the model. If it hits, Yankı stops reflecting and shows human contacts. The model does not get to decide this.

Reflections are produced through Anthropic's Messages API with claude-haiku-4-5-20251001 under a reflection-constrained prompt that must return JSON. The output is schema-checked server side; a malformed answer is discarded rather than shown. Crisis wording is decided by a rule list before the model is called.