# Prompt Generator V2

> Tạo prompt nâng cao theo framework KERNEL (Context, Task, Constraints, Format, Verification), hỗ trợ prompt chaining.

- Skill: `hoangvantuan/prompt-generator-v2` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add hoangvantuan/prompt-generator-v2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/hoangvantuan/prompt-generator-v2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: hoangvantuan (https://skillmd.com/u/hoangvantuan)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/hoangvantuan/prompt-generator-v2

---


# Prompt Generator V2 — KERNEL Framework

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

## KERNEL at a Glance

Each letter is a checkpoint. Chi tiết + ví dụ before/after → [kernel-framework](references/kernel-framework.md).

| Principle | Check | Action if failing |
|---|---|---|
| **K**eep simple | Describe in one sentence? | Split into prompt chain |
| **E**asy to verify | Stranger could verify? | Add measurable criteria |
| **R**eproducible | Works in 30 days? | Remove temporal refs, add versions |
| **N**arrow scope | One deliverable? | Extract goals into separate prompts |
| **E**xplicit constraints | 2-3 "do NOT" rules? | Add negative constraints |
| **L**ogical structure | Context→Task→Constraints→Format? | Restructure |

## Workflow

### Step 0: Determine Mode

| 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 |

### Step 1: Understand Intent

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

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

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

### Step 2: Apply KERNEL

Transform intent into a structured prompt. Run each principle as a mental checklist using the table above. For detailed explanations, consult [kernel-framework](references/kernel-framework.md).

### Step 3: Generate the Prompt

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

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

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

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

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

## 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

### Step 4: Verify with KERNEL Checklist

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

### Step 5: Deliver and Iterate

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?"

## Prompt Chaining

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" →

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

## Failure Modes

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 |