AI大厨
- 发布日期
- 2024年12月19日
- 作者
- Jacob Lloyd —— 项目完成后,在 AI 协助下撰写
- 阅读时长
- 约 14 分钟阅读

简单来说: 一段免费的复制粘贴指令,能把聊天AI变成一个贴心的膳食规划助手。它会建议一周的晚餐菜单,给出食谱,还会写一份已经按超市区域整理好的购物清单。它能帮你省去"今晚吃什么"的纠结,而且不花一分钱。
只需复制粘贴一段系统提示词,就能把聊天模型变成一位膳食规划师:一周的晚餐、对应的食谱,还有一份已经按超市货架分区排好的购物清单。我管他叫ProChef,他还戴着一顶帽子。
tl;dr
- 这是什么: 一段系统提示词("ProChef"),让聊天AI帮你安排一周的餐食。
- 花费: 零——它就是一段提示词。
- 你需要: 任何一个聊天模型都行。我在本地跑Ollama + Open WebUI,不过普通的ChatGPT也一样能用。
- 最终得到: 一周的晚餐计划、带步骤的食谱,以及一份按货架分组的购物清单。
最终成果

你告诉它想要什么,它会提议几道菜,你挑出喜欢的,然后再让它给食谱和购物清单。清单会按生鲜、肉蛋、谷物和常备调料分组返回——这正是能在超市里真正省下时间的部分。
你需要什么
我的环境是本地部署:双卡RTX 3090加一台Threadripper,后端用Ollama,前端用Open WebUI。如果你手头也有48 GB显存,可以原样照抄。
没有也别担心。任何能接受系统提示词的工具都能用,就算你用的工具没有这个选项,像ChatGPT这样把同一段文字粘贴成第一条消息也照样管用。
那张可选的大厨头像,我用的是getimg.ai。没有赞助,没有合作关系,纯粹是我自己在用而已。
准备模型
如果用Ollama,拉取DeepSeek R1-70B。如果"在家跑一个70B模型"听起来像一个大工程,那它确实是——具体步骤我写在了4步本地部署DeepSeek指南里。
如果用ChatGPT或其他托管服务,默认模型就够用。要是服务商刚好有偏食谱风格的模型,用那个也行。
有一项设置很重要:在模型参数里,把上下文窗口设成显存能承受的大小(我这边用DeepSeek R1-70B时大约是5,000个token的上下文——模型本体几乎吃光了全部48 GB,上下文只能捡剩下的边角料)。设好之后保存。
ProChef系统提示词
先坦白一句:这段提示词是我借助AI写的,所以里面混着"seamless user interaction"这种没有人类会真的说出口的措辞。它照样留着,因为它管用。请把方框里的内容一字不差地全部复制,包括结尾那行IMPORTANT——之所以加上它,是因为模型总是悄悄从购物清单里漏掉一些食材。
提示词本身保留英文原文,照抄就行;和它对话时用中文提问,大多数模型也会用中文回复你。
ProChef系统提示词(全部复制)
First: suggest recipes for a meal plan, don't give full recipes or grocery lists until asked.
You are ProChef, a streamlined AI chef assistant designed to simplify meal planning and preparation. Your primary tasks include determining weekly meals based on user preferences, generating detailed ingredient lists with quantities for each selected dish, providing step-by-step cooking directions upon request, and finally creating an organized grocery list when prompted.
Workflow:
1. Meal Determination: Collaborate with the user to establish a meal plan for the week by understanding their dietary restrictions (vegetarian, gluten-free, etc.), budget, time constraints, food preferences, and any specific requests or events.
2. Ingredient Lists & Cooking Directions: Once meals are determined, upon user request, generate precise ingredient lists with quantities for each dish and provide clear, step-by-step cooking directions for preparing the selected meals.
3. Grocery List Generation: When prompted by the user, create a comprehensive grocery list organized by grocery store sections (e.g., produce, dairy, grains). Ensure all necessary ingredients are listed, taking into account standard pantry staples that might not need to be purchased. Allow users to adjust quantities or remove items they already have.Key Specifications:
- User Interaction: Maintain a straightforward and responsive interface for seamless user interaction, ensuring clarity in communication regarding meal plans, ingredient lists, cooking instructions, and grocery lists.
- Accuracy and Completeness: Prioritize the accuracy of information provided. All ingredients must be listed completely to prevent any missing items during shopping or preparation.
- Organizational Efficiency: Organize the grocery list by typical grocery store sections for efficient shopping experiences.Example User Interaction:
User: "I'm ready to make my grocery list."
ProChef: "Here's your organized grocery list. Under 'Produce,' you'll need: 1 head of lettuce, 3 tomatoes... Under 'Dairy'..."By following this streamlined workflow, ProChef ensures that users can efficiently plan, shop for, and prepare their meals with all necessary ingredients accounted for.
IMPORTANT: after generating a grocery list, check the list against the recipes for all ingredients, both implied and called out, to verify all ingredients are accounted for.
贴到哪里:如果你的工具有系统提示词的输入框(Open WebUI在模型设置里),粘进去保存就行。如果没有——ChatGPT也是这种情况——把同样的文字原样贴成新对话的第一条消息,然后照常继续。结果是一样的。
可选:给你的大厨配一张脸
完全没必要的一步,我还是做了。跟一个头像空白的助手聊天,总不如跟一个几何风小厨师聊天有意思,而且顶多绕路五分钟。
我给大厨logo用的getimg.ai设置
模式: AI Image Generator,Essential
模型: Flux.1
提示词: geometric image of a cute chef. Happy and ready to help generate delicious food. he has a mustache and hat which covers his eyes.
负面提示词: Disfigured, cartoon, blurry, nude, background, weapon
生成数量: 一开始每次两张,接近满意后最后一轮拉到10张。

生成,调整提示词,再生成。这个形象大概花了我十几轮提示词修改和差不多100张图。选一张满意的,想放大就放大,下载下来,设成模型的头像。

让它给你安排一周的饭菜
把你的限制条件告诉它,让它来提议。大概是这样:
嘿,我想要下周的晚餐来点大胆的选择,麻烦给我推荐五道能勾起食欲的菜,同时还要健康、有滋有味。请给我两个以上的方案,只要晚餐就行。
它会回复几道候选菜。跟它反复磨合,直到这一周的安排看着顺眼,再要食谱和用量,最后要购物清单。这段提示词是故意设计成"不问就不给食谱"的,这样你不用为了看一眼计划就滚动过一大段文字。
真实案例,未经修改
下面全部都是一次性的回复——第一次回答,没有重试,没有整理。你的结果会不一样。以下的提问和模型回复原本都是用英文进行的,所以这里的截图和引用也保留英文原文。
案例:讨好意大利岳母的假日一周
提出的要求:


它给出的是稳妥的一周安排:周一是意式烤面包片配蔬菜浓汤,接下来是鸡肉青酱意面、茄子帕玛森、奶油培根意面、鸡肉皮卡塔、烤蔬菜配煎鱼,周日则以经典千层面搭配凯撒沙拉收尾。没有猎奇的菜式,每一道都是"认识的菜"——考虑到这次的对象,这个选择是对的。
接着你继续追问具体食谱:

……它照办了,给出了完整步骤。那一大段文字我就不搬过来了——下一个例子里有一份完整的。
案例:两人份、乳糖不耐——附完整输出
提出的要求:"给我一份适合乳糖不耐受人士和她丈夫的每周晚餐计划,食材尽量精简,做法保持简单,并附上按超市货架分类的购物清单。"
模型的回复,一字不改(英文):
Weekly Dinner Meal Plan
Monday: Grilled Salmon with Quinoa Salad
- Ingredients: Salmon fillets, quinoa, bell peppers, broccoli, snap peas, lemon juice.
- Instructions: Grill salmon, cook quinoa, mix with roasted vegetables, dress with lemon juice.
Tuesday: Chicken Stir-Fry with Rice
- Ingredients: Chicken breast, rice, bell peppers, broccoli, snap peas, soy sauce/tamari.
- Instructions: Stir-fry chicken and vegetables in olive oil, serve over rice.
Wednesday: Spaghetti Squash with Ground Turkey Tomato Sauce
- Ingredients: Spaghetti squash, ground turkey, tomato sauce (dairy-free), garlic.
- Instructions: Bake spaghetti squash, cook ground turkey with tomato sauce, combine.
Thursday: Tacos with Ground Beef/Turkey and Veggies
- Ingredients: Corn tortillas, ground beef/turkey, lettuce, tomatoes, avocado, taco seasoning.
- Instructions: Cook ground meat with taco seasoning, serve in tortillas with vegetables.
Friday: Baked Cod with Asparagus and Potatoes
- Ingredients: Cod fillets, asparagus, potatoes, lemon juice, herbs (thyme, rosemary).
- Instructions: Bake cod, roast asparagus and potatoes with olive oil and seasonings.
Saturday: Shrimp Skewers with Peanut Sauce over Couscous
- Ingredients: Shrimp, couscous, peanut butter, garlic.
- Instructions: Grill shrimp skewers, cook couscous, serve with peanut sauce.
Sunday: Turkey Burgers over Greens
- Ingredients: Turkey burgers (or patties), lettuce, tomatoes, onion, avocado.
- Instructions: Cook turkey burgers, serve on a bed of greens with vegetables.
Grocery List by Store Sections
Produce:
- Bell peppers (2)
- Broccoli (1 head)
- Snap peas (1 cup)
- Lettuce (1 head)
- Tomatoes (3)
- Avocado (3)
- Asparagus (1 bunch)
- Potatoes (4 medium)
- Zucchini (1)
- Yellow squash (1)
- Onion (2)
- Cucumber (1)
- Olives (optional)
- Spinach (1 bag)
- Lemons (2)
Proteins:
- Salmon fillets (4 portions)
- Chicken breasts (4 portions)
- Ground turkey (1 lb for spaghetti squash and 1 lb for tacos/ground beef alternative)
- Cod fillets (4 portions)
- Shrimp (1 bag, frozen)
- Turkey burgers (4 patties) or ground turkey to make patties
Grains:
- Quinoa (2 cups)
- Rice (2 cups)
- Corn tortillas (8-10)
- Couscous (2 cups)
Pantry:
- Olive oil
- Coconut oil
- Soy sauce/tamari
- Tomato sauce (dairy-free)
- Taco seasoning
- Peanut butter
- Salt and pepper
- Herbs (oregano, basil, thyme, rosemary)
Miscellaneous:
- Garlic
- Lemon zest or juice
This meal plan ensures minimal ingredients overlap, simple cooking techniques, and no dairy products. Enjoy your week of delicious and easy meals!
注意生鲜那一栏列出了西葫芦、黄南瓜、黄瓜、橄榄和菠菜,而上面任何一道食谱都没提到这些——常备调料里还多了椰子油。就当是白送的蔬菜吧。这正是提示词要求它二次核对清单的原因,也是为什么你自己也该扫一眼的原因。我还用同一个请求跑了一份无麸质版本,结果形状一样——七天无麸质晚餐加一份按货架分类的清单,连高汤都标注了无麸质。
案例:周末约会晚餐菜单
提出的要求:"新任务。帮我安排一个周末给女朋友做饭的完整晚餐计划。菜要够惊艳,每道菜都要有完整做法,再加一份按超市货架分类的购物清单。"
它彻底走起了浪漫路线:周五三文鱼配芒果莎莎酱,周六羊排配蒜香土豆泥,周日红酒汁牛里脊——完整食谱、购物清单,外加一句一本正经的"摆盘要有艺术感"提醒。结尾还留了一句"祝你享受这段烹饪之旅!",一看就知道是谁写的。不过话说回来,这份菜单本身确实相当过硬。
容易踩的坑
- 本地硬件上下文很紧张。 在48 GB显存上跑70B模型,留给我的窗口大约只有5,000个token。一份晚餐计划加食谱再加购物清单,轻轻松松就超了,模型开始忘记对话最前面的内容。每周开新对话,并且赶在内容变长之前就要购物清单。
- 它会漏材料(也会自己加材料)。 提示词末尾的IMPORTANT那行,就是因为购物清单老是漏掉食谱里明显需要的东西才加上的。有帮助,但治不了根。买东西前自己拿清单对一遍食谱。
- 数量值得多留个心眼。 有一份清单给分配了牛肉的塔可要了火鸡肉末。凑合着能去买,但不是圣旨。
- 推理模型会先想后说。 R1回答前会闷头想上一阵子。对膳食规划来说这没关系——别以为它卡死了就行。