what-to-eatlisted

# 吃什么 / What to Eat Give a concrete decision quickly, learn from actual eating behavior, and prevent repetitive suggestions. Treat allergies, dietary restrictions, and explicit exclusions as hard constraints. ## Start every turn 1. Run `python3 scripts/meal_memory.py snapshot` from this skill directory. 2. Read the JSON snapshot before recommending or interpreting numbered feedback. 3. Use the user’s explicit meal or time reference. Otherwise infer breakfast at 07:00–09:30, lunch at 11:30–13:30, and dinner at 17:30–20:00 in the user’s local timezone. If none applies, ask whether this is breakfast, lunch, or dinner. 4. Never claim that profile, history, or feedback was saved unless the matching script command succeeded. The script stores private state locally under `~/.codex/state/what-to-eat/`. Override this in tests with `--state-dir PATH`. ## Handle first use If `profile.completed` is false, ask a lightweight onboarding question. Ask at most two questions per turn; collect safety-critical restrictions first: 1. Allergies, dietary restrictions, and foods the user never eats. 2. Preferred tastes and cuisines. 3. Current goal: healthy/light, weight control, gain energy, satisfying, convenient, economical, comfort, or none. 4. Accepted spice level. 5. Output preference: names only, reasons, pairings, or a simple method. 6. Usual breakfast, lunch, and dinner times, asked together in one low-cost question. Explain that scheduled suggestions are optional. Do not block an i