Prompt Generator V2 — KERNEL Framework

提示词生成器V2 — KERNEL框架

Generate prompts that work on the first try. The KERNEL framework ensures every prompt has a clear goal, verifiable success criteria, and explicit constraints.

生成首次使用即可生效的提示词。KERNEL框架确保每个提示词都具备清晰的目标、可验证的成功标准以及明确的约束条件。

KERNEL at a Glance

KERNEL框架概览

Each letter is a checkpoint. Chi tiết + ví dụ before/after → kernel-framework.

Principle	Check	Action if failing
Keep simple	Describe in one sentence?	Split into prompt chain
Easy to verify	Stranger could verify?	Add measurable criteria
Reproducible	Works in 30 days?	Remove temporal refs, add versions
Narrow scope	One deliverable?	Extract goals into separate prompts
Explicit constraints	2-3 "do NOT" rules?	Add negative constraints
Logical structure	Context→Task→Constraints→Format?	Restructure

每个字母对应一项检查标准。详情+前后示例→kernel-framework。

原则	检查标准	未达标时的处理方式
K简洁性（Keep simple）	能否用一句话描述核心目标？	拆分为提示词链
E可验证性（Easy to verify）	陌生人能否验证结果是否符合要求？	添加可衡量的判定标准
R可复现性（Reproducible）	30天后仍能生效？	移除时效性参考，添加版本信息
N窄范围（Narrow scope）	仅对应一项交付成果？	将多个目标拆分为独立的提示词
E明确约束（Explicit constraints）	是否包含2-3条“禁止”规则？	添加负面约束条件
L逻辑结构（Logical structure）	是否遵循“上下文→任务→约束→格式”的结构？	重新调整结构

Workflow

工作流程

Step 0: Determine Mode

步骤0：确定模式

User input	Mode	Action
Vague request ("help me write a prompt for X")	Create	Go to Step 1
Existing prompt provided	Improve	Run KERNEL checklist against the prompt, diagnose which principles fail, fix targeted. Skip to Step 2

用户输入	模式	操作
模糊需求（如“帮我写一个针对X的提示词”）	创建模式	进入步骤1
提供现有提示词	优化模式	用KERNEL检查清单评估该提示词，诊断未达标的原则并针对性修复，直接进入步骤2

Step 1: Understand Intent

步骤1：明确需求意图

Extract or ask (max 3 questions — skip if the request already answers them):

What's the single goal? — If multiple goals detected, suggest splitting into a prompt chain
What does success look like? — Specific, verifiable criteria (numbers, formats, concrete deliverables)
What should it NOT do? — Constraints and exclusions

If the user provides a vague request, propose a draft immediately and iterate — action beats interrogation.

提取信息或询问用户（最多3个问题——若需求已明确则跳过）：

核心目标是什么？——若检测到多个目标，建议拆分为提示词链
成功的标准是什么？——具体、可验证的判定条件（如数字、格式、具体交付成果）
需要禁止哪些行为？——约束条件与排除项

若用户提供的需求模糊，应立即生成草稿并迭代——行动优先于追问。

Step 2: Apply KERNEL

步骤2：应用KERNEL框架

Transform intent into a structured prompt. Run each principle as a mental checklist using the table above. For detailed explanations, consult kernel-framework.

将需求意图转化为结构化提示词。以上述表格为参考，逐一核对每项原则。如需详细说明，请查阅kernel-framework。

Step 3: Generate the Prompt

步骤3：生成提示词

Use this structure. Include only relevant sections — omit what doesn't apply:

undefined

使用以下结构，仅保留相关部分，省略不适用的内容：

undefined

Context

[Background information the AI needs. Keep minimal — only what's necessary to understand the task. Include domain, audience, and relevant technical context.]

[AI所需的背景信息。尽量精简——仅保留理解任务必需的内容，包括领域、受众及相关技术背景。]

Task

[One clear, specific goal. Start with an action verb. This is the single sentence that passes the K-test.]

[一个清晰、具体的目标。以动作动词开头。这是符合K原则的单句描述。]

Constraints

[What to do — specific, measurable behaviors]
Do NOT [negative constraint 1]
Do NOT [negative constraint 2]
[Additional bounds: length, format, libraries, scope limits]

[需要执行的具体、可衡量的行为]
Do NOT [禁止行为1]
Do NOT [禁止行为2]
[额外限制：长度、格式、使用的库、范围限制]

Output Format

[Exact structure of the expected output. Include: format (markdown, JSON, code), length bounds, sections/headers if applicable, delimiters.]

[预期输出的精确结构。包括：格式（markdown、JSON、代码）、长度限制、适用的章节/标题、分隔符。]

Verification

[How to check success — specific criteria that make the E-principle concrete. Think: "I'll know this worked when..."]


**Optional sections** (include when they add value):

- **Examples** — When output quality depends on seeing patterns (2-3 examples: basic + edge case)
- **Input** — When the prompt processes structured data (describe format, required fields)
- **Chain** — When the task was split, show how prompts connect

[如何验证成功——使E原则落地的具体判定标准。思考：“当出现XX情况时，我就知道这生效了”]


**可选章节**（仅在能提升效果时添加）：

- **示例**——当输出质量依赖于参考模式时（2-3个示例：基础案例+边缘案例）
- **输入**——当提示词需要处理结构化数据时（描述格式、必填字段）
- **链结构**——当任务被拆分时，说明各提示词之间的关联方式

Step 4: Verify with KERNEL Checklist

步骤4：用KERNEL检查清单验证

Before delivering, run this self-review:

K: Can I describe this prompt's goal in one sentence?
E: At least 2 measurable success criteria?
R: No temporal references, no version-ambiguous terms?
N: Exactly one deliverable per prompt?
E: At least 2 explicit "do NOT" constraints?
L: Follows Context → Task → Constraints → Format structure?
No vague virtue words ("good", "helpful", "detailed") without concrete definition
No contradictions (e.g., "be concise" + "cover everything")
All implicit assumptions made explicit

在交付前，执行以下自我检查：

K：能否用一句话描述该提示词的核心目标？
E：是否包含至少2条可衡量的成功标准？
R：无时效性参考内容，无版本模糊的术语？
N：每个提示词仅对应一项交付成果？
E：是否包含至少2条明确的“禁止”约束？
L：是否遵循“上下文→任务→约束→格式”的结构？
无未给出明确定义的模糊形容词（如“优秀”、“有用”、“详细”）
无矛盾要求（如“保持简洁”+“涵盖所有内容”）
所有隐含假设均已明确说明

Step 5: Deliver and Iterate

步骤5：交付与迭代

Present the prompt in a clean code block. If the original request was complex and got split:

Show each prompt in the chain, numbered
Explain how outputs feed into subsequent prompts
Suggest which prompts can run in parallel vs sequential

Always offer: "Want me to adjust the constraints, add examples, or split this differently?"

将提示词放在清晰的代码块中展示。若原需求复杂且已拆分：

按编号展示链中的每个提示词
说明前一个提示词的输出如何作为下一个的输入
建议哪些提示词可并行执行，哪些需按顺序执行

始终主动询问：“需要我调整约束条件、添加示例或换一种拆分方式吗？”

Prompt Chaining

提示词链

When a task is too complex for one prompt (fails N-principle), decompose into a chain. Each link:

Has a single clear goal (passes all KERNEL checks independently)
Produces output that feeds cleanly into the next prompt
Can be verified independently before moving to the next step

Pattern: Task → subtask analysis → ordered chain with data flow

Example: "Build a REST API" →

Design data models (output: schema)
Generate endpoint specifications (input: schema → output: OpenAPI spec)
Implement endpoints (input: OpenAPI spec → output: code)
Write tests (input: code + spec → output: test suite)

当单个提示词无法处理复杂任务（未通过N原则）时，将其拆分为提示词链。链中的每个环节需满足：

具备单一清晰的目标（可独立通过所有KERNEL原则检查）
生成的输出可直接作为下一个提示词的输入
在进入下一个环节前可独立验证结果

模式：任务→子任务分析→带数据流的有序链

示例：“构建一个REST API”→

设计数据模型（输出：Schema）
生成接口规格（输入：Schema → 输出：OpenAPI规范）
实现接口（输入：OpenAPI规范 → 输出：代码）
编写测试用例（输入：代码+规范 → 输出：测试套件）

Failure Modes

常见错误模式

Các lỗi phổ biến cần nhận diện và tránh khi generate prompt:

Failure mode	Dấu hiệu	Sửa
Prompt quá chung	Không constraint, output có thể là bất kỳ thứ gì	Thêm scope, format, length bounds
Over-engineering	Prompt dài hơn output mong đợi, quá nhiều rules	Cắt constraints không ảnh hưởng output quality
Constraint mâu thuẫn	"Be concise" + "Cover everything thoroughly"	Chọn 1, bỏ kia, hoặc chia scope
Vague virtue stacking	"Good", "helpful", "engaging", "detailed" liên tiếp	Thay bằng criteria cụ thể, đo được
Temporal drift	"Current", "latest", "recent" không pin version	Pin version/date cụ thể
Missing audience	Prompt không nói cho ai → tone/depth không phù hợp	Thêm audience + expertise level
Format ambiguity	Không nói rõ output format → AI tự chọn	Thêm Output Format section tường minh

生成提示词时需识别并避免的常见错误：

错误模式	迹象	修复方法
提示词过于宽泛	无约束条件，输出内容无边界	添加范围、格式、长度限制
过度设计	提示词比预期输出还长，规则过多	删除不影响输出质量的约束条件
约束条件矛盾	同时要求“保持简洁”和“全面涵盖所有内容”	二选一，或拆分任务范围
模糊形容词堆砌	连续使用“优秀”、“有用”、“吸引人”、“详细”等模糊词汇	替换为具体、可衡量的判定标准
时效性偏差	使用“当前”、“最新”、“近期”等未绑定版本的词汇	绑定具体版本或日期
缺失受众信息	提示词未说明面向对象，导致语气/深度不符	添加受众及专业水平信息
格式模糊	未明确输出格式，由AI自行选择	添加明确的“输出格式”章节

prompt-generator-v2

Original

Translation

Prompt Generator V2 — KERNEL Framework

提示词生成器V2 — KERNEL框架

KERNEL at a Glance

KERNEL框架概览

Workflow

工作流程

Step 0: Determine Mode

步骤0：确定模式

Step 1: Understand Intent

步骤1：明确需求意图

Step 2: Apply KERNEL

步骤2：应用KERNEL框架

Step 3: Generate the Prompt

步骤3：生成提示词

Context

Context

Task

Task

Constraints

Constraints

Output Format

Output Format

Verification

Verification

Step 4: Verify with KERNEL Checklist

步骤4：用KERNEL检查清单验证

Step 5: Deliver and Iterate

步骤5：交付与迭代

Prompt Chaining

提示词链

Failure Modes

常见错误模式