上周有个学员找我，说他的英语批改助手"抽风"：同一段代码，上午批改返回三行规整的"错在哪—为什么—怎么改"，下午就变成一篇热情洋溢的小作文，晚上干脆给他打了个分还附赠评语。
代码一行没动。他怀疑是模型坏了，一度想写个定时重启脚本来"修"它。
我看了他的代码就明白了：system prompt 是一句"你是一个英语老师，帮助学生批改作文"。这一句话，等于什么都没说。模型每次自由发挥，风格当然像抽签。
这一篇我们就把这件事彻底讲透：提示词到底是怎么决定回答的，以及怎么把它变成一个可测试、可版本管理的工程对象，而不是散落在代码里的一句祈祷词。
一、先扫盲：为什么 system prompt 有这么大权力
1. 三种角色，各管各的
上一篇讲过，一次请求里的消息分三种角色：
- system：定身份和规矩。用户看不见它，但它对模型输出的影响最大。
- user：用户真正说的话。
- assistant：模型之前的回答，用于多轮对话。
一个容易被忽略的事实：模型对 system 的服从度远高于 user。 你在 user 消息里说"请用表格回答"，模型可能理会也可能不理；同样的要求写进 system，基本每次都执行。所以身份、输出格式、语气这些"每次都该生效"的规矩，全部放 system。
2. 输出不稳定，是因为你只给了目标没给规格
你说"批改这篇作文"，模型不知道你要：一句话点评？逐句批注？打分加清单？它只能猜。猜对猜错，全看当天的心情（准确说是采样随机性）。
解决思路有两种，一种贵一种便宜：
- 贵的思路：让模型输出 JSON，程序解析后强制重渲染。这是下一篇的主题，适合格式必须严格的场景。
- 便宜的思路：给两三个示例，让模型照着写。 这就是少样本（few-shot），今天讲。
3. temperature：同一个提示词，为什么两次回答不一样
模型生成每个字都是一个概率分布，temperature 控制从分布里怎么挑：
批改助手要的是"同一篇文章每次给出一致的批改"，把 temperature 压到 0.2 左右。很多"模型抽风"问题，一半是 prompt 没写清，另一半就是 temperature 没调。
4. 思维链：让模型先打草稿再交卷
模型是逐字生成的。让它直接给结论，相当于默算；让它先写中间步骤，等于把草稿摊在桌面上，每一步都能踩住前一步的正确结果。
对语法批改这种"先找成分、再套规则"的活，思维链提升明显。但它会让输出变长、token 变多，只留给复杂任务，别逢题就挂。
二、工程设计：把提示词从"字符串"升级成"对象"
第 1 篇的代码里，system prompt 是写死在函数里的字符串。这一篇要改掉这个坏习惯。
新工程结构（settings.py 和 llm.py 与第 1 篇完全一致，原样复用）：
02-Prompt工程/
├── src/
│ ├── settings.py # 复用第 1 篇
│ ├── llm.py # 复用第 1 篇
│ ├── prompts.py # 人格字典 + 任务模板
│ ├── fewshot.py # 少样本示例
│ ├── cot.py # 思维链三档
│ └── ask.py # 统一入口
├── tests/ # 27 个用例，离线全绿
└── demo.py
核心思想一句话：人格管"怎么说"，任务管"做什么"，示例管"输出长什么样"，思维链管"解题步骤"——四样东西拆开，才能自由组合。
三、人格与任务：两段 system，分开管（prompts.py）
先看数据：
PERSONAS: dict[str, str] = {
"strict": """你是「启明英语」的资深语法讲师，以严格著称。
要求：
1. 直接指出错误，不铺垫、不夸奖
2. 每个错误给出：错在哪 → 为什么错 → 正确写法
3. 回答控制在 3 句以内
""",
"encouraging": """你是「启明英语」的陪伴型助教，面对的是容易灰心的初学者。
要求：
1. 先肯定做对的部分，再指出要改的地方
2. 纠正时给出完整正确句，不要用术语
3. 语气像同班同学，不要用"建议您"这类官话
""",
"examiner": """你是「启明英语」的阅卷官，按雅思写作评分标准打分。
要求：
1. 先给总分（0-9，可以带 .5），再给四项小分
2. 每项小分后必须跟一句扣分原因
3. 最后给一条最优先的改进建议
""",
}
注意这三段文案的共同点：都有编号要求、都有可验证的格式约束（"3 句以内""必须跟一句原因"）。 "你是一个专业的英语老师"这种话在要求栏里是零信息量的——专业不专业，模型自己说了不算。
任务模板带变量：
TASK_TEMPLATES: dict[str, str] = {
"grammar": "请批改下面这句英语的语法，目标对象是 {level} 水平的学生。",
"polish": "请把下面这句英语改得更自然，目标对象是 {level} 水平的学生。",
"quiz": "请就下面这个知识点出一道题，难度匹配 {level} 水平。",
}
组装用 LangChain 的 ChatPromptTemplate，这是本篇第一个"用框架"的点：
def build_prompt(
persona: str = DEFAULT_PERSONA,
task: str = DEFAULT_TASK,
) -> ChatPromptTemplate:
"""拼出一个完整的对话模板。
结构是固定的三段：
system 人格 → 告诉模型"你是谁、怎么说话"
system 任务 → 告诉模型"这次要做什么"
human 提问 → 学生真正输入的内容
"""
return ChatPromptTemplate.from_messages(
[
("system", get_persona(persona)),
("system", get_task(task)),
("human", "{question}"),
]
)
这里我刻意用了两个 system 而不是合并成一段。合并当然也能跑，但拆开后"换人格不动任务、换任务不动人格"，做产品时你会感谢这个决定。
用框架还有个隐藏好处：{level} 这种变量少传了会直接抛 KeyError，而不是静默生成一句带花括号的残缺文本。字符串拼接时代这种事故一查一晚上，框架替你把这类错误挡在了上线前。
四、少样本：把输出规格"演示"一遍（fewshot.py）
ESSAY_EXAMPLES: list[dict[str, str]] = [
{
"input": "I go to the park yesterday.",
"output": (
"错误：时态\n"
"原文：I go to the park yesterday.\n"
"改正：I went to the park yesterday.\n"
"理由：yesterday 是明确的过去时间，动词要用过去式。"
),
},
# …… 共 3 条
]
组装时用 FewShotChatMessagePromptTemplate：
EXAMPLE_PROMPT = ChatPromptTemplate.from_messages(
[
("human", "{input}"),
("ai", "{output}"),
]
)
FEW_SHOT_PROMPT = FewShotChatMessagePromptTemplate(
example_prompt=EXAMPLE_PROMPT,
examples=ESSAY_EXAMPLES,
)
def build_fewshot_prompt(system_text: str) -> ChatPromptTemplate:
"""顺序很关键：人格 → 示例 → 真正的提问。"""
return ChatPromptTemplate.from_messages(
[
("system", system_text),
FEW_SHOT_PROMPT,
("human", "{question}"),
]
)
两个必须记住的坑：
- 示例质量决定输出上限。 示例里的 assistant 回答会被当成模板照抄，示例写得敷衍，输出就敷衍。
- 提问必须排在最后。 我在代码注释里写了原因：示例放最后，模型会把最后一条示例输出当成"上一轮对话"，接着往下续写。
这张图就是挂了 3 条示例后，真正发给模型的完整消息列表——8 条。你可以打印出来逐条核对，我强烈建议你养成这个习惯：提示词调试的第一步不是盯着回答猜，而是把喂进去的原文看一遍。
五、思维链：三档强度，按需取用（cot.py）
COT_LEVELS: dict[str, str] = {
"light": "先说结论，再用一句话解释原因。",
"standard": """按下面三步来：
1. 找出句子里的时间状语和谓语动词
2. 判断时态是否正确
3. 给出改正后的句子
""",
"heavy": """请严格按下面的步骤拆解，每一步单独一行：
第 1 步 · 成分分析：主语是____，谓语是____，时间状语是____
第 2 步 · 时态判断：这个时间状语要求用____时态
第 3 步 · 问题定位：原文的谓语动词是____，与要求____（一致/不一致）
第 4 步 · 改写句子：____
第 5 步 · 中文解释：____
""",
}
组装时我把思维链挂在 human 消息而不是 system：
def build_cot_prompt(persona: str, cot_level: str = DEFAULT_COT) -> ChatPromptTemplate:
"""思维链引导语放 human 里：system 是长期性格，
思维链是本次任务的执行步骤，混在一起会让想换人格的人连带改步骤。"""
return ChatPromptTemplate.from_messages(
[
("system", persona),
("human", "{cot}\n\n句子：{question}"),
]
).partial(cot=get_cot(cot_level))
.partial() 是模板的预填变量功能：先钉死 cot，剩下的 question 调用时再传。
分工原则记住一句话：system 放不变的东西，human 放每次会变的东西。 人格三周不换，题型每天变——放错了位置，改起来会互相牵连。
六、统一入口：四个开关自由组合（ask.py）
def ask(
question: str,
llm: BaseChatModel,
*,
persona: str = "strict",
task: str = "grammar",
level: str = "初中",
fewshot: bool = False,
cot: str | None = None,
) -> str:
persona_text = get_persona(persona)
if fewshot:
prompt = build_fewshot_prompt(persona_text)
messages = prompt.format_messages(question=question)
elif cot:
prompt = build_cot_prompt(persona_text, cot_level=cot)
messages = prompt.format_messages(question=question)
else:
prompt = build_prompt(persona=persona, task=task)
messages = prompt.format_messages(question=question, level=level)
response = llm.invoke(messages)
return response.content
调用方从此只写一行：ask(句子, llm, persona="examiner", fewshot=True)。
七、跑起来：真实输出
python demo.py，同一个句子 He go to school by bus every day.，三种人格渲染出的提示词对比（节选 strict）：
──────── 同一个问题，三种人格（问题：He go to school by bus every day.） ────────
【strict】
[SystemMessage] 你是「启明英语」的资深语法讲师，以严格著称。
要求：
1. 直接指出错误，不铺垫、不夸奖
2. 每个错误给出：错在哪 → 为什么错 → 正确写法
3. 回答控制在 3 句以内
[SystemMessage] 请批改下面这句英语的语法，目标对象是 初中 水平的学生。
[HumanMessage] He go to school by bus every day.
离线链路自检（预置答案，验证四种人格真的走了不同分支）：
[strict ] 时态错误：主语 He 是第三人称单数，every day 要求一般现在时，应改为 He goes。
[encouraging] 这句有点小问题哦，'He go' 要改成 'He goes'，其他都对啦！
[examiner ] 总分 5.0。语法项扣分：主谓不一致（He go → He goes）。
测试结果：
tests/test_ask.py::test_ask_returns_model_text PASSED
tests/test_ask.py::test_persona_text_is_really_sent PASSED
tests/test_ask.py::test_task_level_reaches_the_request PASSED
tests/test_ask.py::test_fewshot_sends_more_messages PASSED
tests/test_ask.py::test_fewshot_includes_example_content PASSED
tests/test_ask.py::test_cot_adds_guidance_to_human_message PASSED
tests/test_ask.py::test_fewshot_takes_priority_over_cot PASSED
tests/test_cot.py::test_all_levels_available PASSED
tests/test_cot.py::test_unknown_level_raises PASSED
tests/test_cot.py::test_heavy_has_more_steps_than_light PASSED
tests/test_cot.py::test_cot_text_is_sent_to_model PASSED
tests/test_cot.py::test_cot_sits_in_human_not_system PASSED
tests/test_cot.py::test_different_levels_give_different_prompts PASSED
tests/test_fewshot.py::test_examples_are_configured PASSED
tests/test_fewshot.py::test_each_example_has_input_and_output PASSED
tests/test_fewshot.py::test_fewshot_message_count PASSED
tests/test_fewshot.py::test_question_is_the_last_message PASSED
tests/test_fewshot.py::test_examples_alternate_human_ai PASSED
tests/test_fewshot.py::test_example_content_reaches_model PASSED
tests/test_prompts.py::test_default_persona_exists PASSED
tests/test_prompts.py::test_unknown_persona_raises PASSED
tests/test_prompts.py::test_unknown_task_raises PASSED
tests/test_prompts.py::test_render_has_two_system_and_one_human PASSED
tests/test_prompts.py::test_level_variable_is_interpolated PASSED
tests/test_prompts.py::test_missing_variable_raises PASSED
tests/test_prompts.py::test_different_personas_give_different_text PASSED
tests/test_prompts.py::test_render_output_is_readable PASSED
27 passed in 0.32s
27 个用例全绿，全程离线。 这组测试里我最想让你注意到 test_persona_text_is_really_sent：它用一个会记录入参的假模型，断言"阅卷官"三个字真的出现在了发给模型的请求里。
为什么这条测试重要？因为提示词开发里最阴险的 bug 就是你以为配了，实际没生效——人格名字拼错、字典取到默认值、变量传了个 None，程序都不报错，回答却悄悄走样。有了记录型假模型，这类问题一测一个准。
八、踩坑清单
- 人格名字写错却安静地用了默认值：get_persona() 刻意不静默兜底，错了就抛 KeyError。宁可当场合身，不要上线后走样。
- 示例排在提问后面：模型把示例输出当上文续写。本工程 build_fewshot_prompt 已固定顺序，别手痒调整。
- 示例写得敷衍：示例是模板，不是参考。示例回答的质量就是输出质量的天花板。
- 思维链逢题就挂：输出变长、token 变多，简单释义场景纯属浪费。只给复杂任务挂。
- 把输出格式约束写在 user 消息里：模型时理时不理。格式要求放 system，一次写清。
- pytest 报 No usable temporary directory：本机安全策略限制临时文件所致，加 --capture=no -p no:cacheprovider 关闭输出捕获即可，与代码无关。
九、本篇收尾
现在你的批改助手有了三种可切换的人格、稳定的输出结构、按需启用的解题步骤，并且每一层都有测试看着。那个"说话像换了个人"的问题，从玄学变成了一组可以打印、可以断言、可以版本管理的对象。
下一篇解决一个更烧钱的问题：聊十轮就"失忆"，账单还翻倍？ 我们会算清 token 这笔账，写出第一个带裁剪策略的会话管理器——你将亲眼看到"多轮对话"是怎么从免费变成账单的。
最后一句：模型不抽风，抽风的是提示词——你以为你说了要求，其实你只是许了个愿。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Last week, a student came to me and said his English-language rephrasing assistant, "Scrambling": "What's wrong with the same code, changing it back to three lines in the morning?"
The code line is intact. He suspected that the model was broken, and he tried to write a scheduled reset.
I read his code, and I figured out: "System prompt is a sentence that says, "You're an English teacher, helping the students change the script." That is nothing. Every time the model is free, the style is certainly like drawing lots.
And this is the one where we can tell the whole story: how exactly prompt decided to answer it, and how to turn it into a testable, versionable project subject, rather than a prayer line scattered in the code.
I. Literacy first: Why is system prompt so powerful?
1. Three roles, one for each.
As stated in the previous article, there are three kinds of messages in a request:
- System: Identity and rules. Users cannot see it, but it has the greatest impact on model output.
-user: The user really said that.
- Alistant: The answer before the model is used for multiple rounds of dialogue.
An easy-to-neglected fact: Models are much more obedient to systems than user. You said, "Please answer in a form" in our source, the model may or may not; the same requirement is written in system, almost every time. So the rules of identity, output format, tone, "it should come into effect every time" are put in the system.
The output was unstable because you gave the target no specifications.
You say, "Crucify this essay," and the model doesn't know you want to say, "Assumption by sentence? It can only guess. And guess right and wrong, depending on the mood of the day (sampling randomly, to be precise).
There are two kinds of solutions, one priceless:
- Your idea: let the model out of JSON, force retrofitting after the program has been analyzed. This is the subject of the next article, and the appropriate format must be rigorous.
- Cheap thinking: give two or three examples of how the model works. That's less sample, speaking today.
3. Temperature: the same prompt, why two different answers
Models generate a probability distribution for every word, how to choose from the distribution:
The batching assistant wants "the same article every time it gives the same rephrasing" to press the temperature around 0.2. There's a lot of "model twitching" problems, and half of it's prompt, and the other half is temperature.
4. Thinking chain: let the model draft before the paper is submitted
Models are created word for word. Making it come directly to conclusions is tantamount to a silent calculation; allowing it to write intermediate steps is tantamount to placing drafts on the table and stepping on the right results of each step.
The grammatical modification of this "first-to-first-to-first-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one-to-one" exercise is a remarkable rise in the thought chain. But it makes the output longer and the token more, leaving it only to complex tasks, and not to be confused.
Engineering Design: Upgrading prompt from "string" to "object"
In the code of Part 1, system policy is a string that dies in a function. This is about changing this bad habit.
New Project Structures (settings.py and llm.py are fully consistent with Title 1 and are duplicated):
02-Prompt Project/
├── src/
{\cHFFFFFF}{\cH00FFFF} {\cHFFFFFF}{\cH00FFFF} Page {\cHFFFFFF}
│llm.py # Reused Part 1
│ - prompts.py # Personal Dictionary + Task Template
Z Z Z Z Z
I'm thinking - cot.py
{\cHFFFFFF}{\cH00FFFF} {\cHFFFFFF}{\cH00FFFF}
offline All Green
└── demo.py
The core thought phrase: "How" and "what" and "what" and "output" and "the thought chain" and "steps" -- all four things to be able to form freely.
Personality and mission: two systems, separate tubes (prompts.py)
See data first:
PERSONAS: dict[str, str] = {
"strict": """You're a senior grammar lecturer for the Enlightenment of English.
Requests:
1. Directly identifying errors without setting aside or exaggerating
Each error gives: where is wrong?
3. Answer within 3 sentence
""",
"Encouraging": ""You're the assistant assistant to the Enlightened English language, and you're faced with an easily disheartening beginner.
Requests:
1. Make sure the right part is done before pointing out where to change
2. Complete correct sentence when corrected, not term
3. Speak like a classmate. Don't use words like "advisory."
""",
"examiner": ""You're the confidant of the English language.
Requests:
1. Total points first (0-9 with 5) and then four small min
2. Every fraction must be followed by a deduction
3. Final recommendations for improvement with the highest priority
""",
}
Attention is drawn to the common denominator of these three paragraphs: they all have a numbering requirement and are subject to a verifiable format ( "must follow" in the third sentence). "You're a professional English teacher" is a zero-information column -- not a professional, not a model.
Task template with variable:
TASK_TEMPLATES: dict[str, str] = {
"grammar": "Please change the English syntax below to target students at {level} level. " I don't know.
"Polish": "Please make the following English more natural and target students at {level} levels. " I don't know.
"quiz": "Please click on the following knowledge to match the {level} level. " I don't know.
}
This is the first "Framework" point of this article:
def build_prompt(
persona: str = DEFAULT_PERSONA,
task: str = DEFAULT_TASK,
) -> ChatPromptTemplate:
"""spelled a complete dialogue template.
The structure is three fixed segments:
Tell the model who you are and how you speak.
You know what, system mission.
Human, ask questions.
"""
return ChatPromptTemplate.from_messages(
[
("system", get_persona(persona)),
("system", get_task(task)),
("human", "{question}"),
]
)
I've deliberately used two systems here instead of merging them into one. The merger can run, of course, but you'll be grateful for that decision when you open it up.
There's also a hidden advantage in using the frame: {level} It's a variable that doesn't pass directly to KeyError, instead of quietly generating a bracketed missing text. The frame kept such errors in front of the line for you as soon as you looked into such accidents in the age of stringing.
iv. Less sample: "Fewshot.py"
ESSAY_EXAMPLES: list[dict[str, str]] = [
{
"input": "I go to the park yesterday.",
"output": (
Error: Time\n
"I go to the park yesterday."
"I went to the park yesterday."
Reason: yessterday is the exact time of the past, the verb must be the past.
),
},
# 3 articles
]
FewShotChatMessagePromptTemplate:
EXAMPLE_PROMPT = ChatPromptTemplate.from_messages(
[
("human", "{input}"),
("ai", "{output}"),
]
)
FEW_SHOT_PROMPT = FewShotChatMessagePromptTemplate(
example_prompt=EXAMPLE_PROMPT,
examples=ESSAY_EXAMPLES,
)
def build_fewshot_prompt(system_text: str) -> ChatPromptTemplate:
""Sequence is critical: personality, example, real question."
return ChatPromptTemplate.from_messages(
[
("system", system_text),
FEW_SHOT_PROMPT,
("human", "{question}"),
]
)
Two pits to remember:
- Example quality determines the output ceiling. The example will be copied as a template, the example will be perfunctory and the output will be perfunctory.
- Questions must be at the end. I wrote the reason in the code note: at the end of the example, the model would take the last example output as "the last round of conversations" and continue down.
This is the complete list of messages that are actually sent to the model - 8 - after 3 examples are hung. You can print it out one by one, and I strongly recommend that you develop this habit: The first step in debugging prompt was not to stare at the answer but to read the original text that had been fed.
V. CUTS OF THINKING: THROUGHT THROUGHS, ON NEEDS (cot.py)
COT_LEVELS: dict[str, str] = {
"light": "Conclusion first, then one sentence to explain why. " I don't know.
"Standard": "Stand" by three steps:
1. Identification of temporal and verbs in sentences
2. Determining whether the times are correct
3. Giving corrected sentences
""",
"heavy": ""Please dismantle it strictly according to the following steps, each step in a separate row:
Step 1 • Component analysis: the principal language is, the term is, the temporal term is
Step 2 • Timestyle judgement: this timestyle requires a -temporal
Step 3 Problem positioning: the verb in the original language is, which is not consistent with the requirements (inconsistent/inconsistent)
Step 4. Reword sentence:
Step 5 Explanation in Chinese:
""",
}
I put my head on a human message instead of system:
def build_cot_prompt(persona: str, cot_level: str = DEFAULT_COT) -> ChatPromptTemplate:
""The mind chain guide" in human: system is long-term character,
The chain of thought is a step in the implementation of this mission, and a combination of it will lead to a change of character.
return ChatPromptTemplate.from_messages(
[
("system", persona),
("human", "cot" )
]
).partial(cot=get_cot(cot_level))
.partial() is the prefilled variable function of the template: the cot first, the rest of the question calls.
The principle of division of labour remembers one thing: system put the same thing, human put the same thing every time it changes. The personality changes for three weeks, the profile changes every day — in the wrong place, and changes are linked to each other.
VI. Unified entrance: free combination of four switches (ask.py)
def ask(
question: str,
llm: BaseChatModel,
*,
persona: str = "strict",
task: str = "grammar",
Level: str = "lower middle school,"
fewshot: bool = False,
cot: str | None = None,
) -> str:
persona_text = get_persona(persona)
if fewshot:
prompt = build_fewshot_prompt(persona_text)
messages = prompt.format_messages(question=question)
elif cot:
prompt = build_cot_prompt(persona_text, cot_level=cot)
messages = prompt.format_messages(question=question)
else:
prompt = build_prompt(persona=persona, task=task)
messages = prompt.format_messages(question=question, level=level)
response = llm.invoke(messages)
return response.content
The caller will write only one line from this point: ask.
VII. RUN UP: REAL OUTCOME
He go to school by bus every day.
He goes to school by bus every day.
【strict】
[SystemMessage] You're a senior grammatical lecturer for the Enlightenment of English.
Requests:
1. Directly identifying errors without setting aside or exaggerating
Each error gives: where is wrong?
3. Answer within 3 sentence
[SystemMessage] Please revise the English syntax below to target students at lower secondary level.
[HumanMessage] He go to school by bus every day.
offline chain self-censorship (preset answers to verify that the four personalities have actually taken different branches):
[strict] Times error: He is a third-person singular, every day requires that normally the time be changed to He goes.
[encourting] That's a little bit of a problem, 'he go'.
[examiner] Total score 5.0. Syntax: He go → He goes.
Test results:
tests/test_ask.py::test_ask_returns_model_text PASSED
tests/test_ask.py::test_persona_text_is_really_sent PASSED
tests/test_ask.py::test_task_level_reaches_the_request PASSED
tests/test_ask.py::test_fewshot_sends_more_messages PASSED
tests/test_ask.py::test_fewshot_includes_example_content PASSED
tests/test_ask.py::test_cot_adds_guidance_to_human_message PASSED
tests/test_ask.py::test_fewshot_takes_priority_over_cot PASSED
tests/test_cot.py::test_all_levels_available PASSED
tests/test_cot.py::test_unknown_level_raises PASSED
tests/test_cot.py::test_heavy_has_more_steps_than_light PASSED
tests/test_cot.py::test_cot_text_is_sent_to_model PASSED
tests/test_cot.py::test_cot_sits_in_human_not_system PASSED
tests/test_cot.py::test_different_levels_give_different_prompts PASSED
tests/test_fewshot.py::test_examples_are_configured PASSED
tests/test_fewshot.py::test_each_example_has_input_and_output PASSED
tests/test_fewshot.py::test_fewshot_message_count PASSED
tests/test_fewshot.py::test_question_is_the_last_message PASSED
tests/test_fewshot.py::test_examples_alternate_human_ai PASSED
tests/test_fewshot.py::test_example_content_reaches_model PASSED
tests/test_prompts.py::test_default_persona_exists PASSED
tests/test_prompts.py::test_unknown_persona_raises PASSED
tests/test_prompts.py::test_unknown_task_raises PASSED
tests/test_prompts.py::test_render_has_two_system_and_one_human PASSED
tests/test_prompts.py::test_level_variable_is_interpolated PASSED
tests/test_prompts.py::test_missing_variable_raises PASSED
tests/test_prompts.py::test_different_personas_give_different_text PASSED
tests/test_prompts.py::test_render_output_is_readable PASSED
27 passed in 0.32s
27 Full green, full offline. I'd like to draw your attention to this set of tests: It uses a fake model that records participation, asserting that the word "diplomar" actually appears in the request for the model.
Why is this test important? Because the most insidious bug in the development of prompt is that you thought it was a match, and it didn't work -- you misspelled the personality, you got the default value in the dictionary, you passed the variable, you didn't say anything wrong, you answered quietly. Documented false models are available, and such problems are measured.
VIII. List of pedestals
- Wrong personal name but silently use the default value: Get persona() deliberately play it quiet, throw KeyError if wrong. I'd rather be present than on the line.
- The examples are followed by questions: the model continues the examples as above. This project built fewshot prompt has been set in order, not to tickle.
- Examples are perfunctory: examples are templates, not references. The quality of the sample answer is the ceiling of the output quality.
- The chain of thought goes hand in hand: the output is longer, the token is more numerous, the simple interpretation is a waste. Only for complex tasks.
- Write output format in user message: Models don't matter. Format requires you to put a system at once.
- Pytest Report No usable security policy limits to temporary file, plus --capture=no-p:cacheprovider closes the output capture, not code.
IX. CONCLUSION
Now your batching assistant has three transposable personalities, a stable output structure, a needs-based solution, and each floor is tested. The question of "spoken like a change of person" went from genre to a group of objects that could be printed, acclaimable, amenable.
The next one's more expensive: 10 rounds of amnesia, double the bill? We'll figure out this token account and write the first session manager with a tailoring strategy -- you'll see for yourself how the "multicycle conversation" went from free to bill.
The last sentence: The model is not windy, it's prompt -- you thought you asked, but you made a wish.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.