SKILL.md
Prompt Generator
Create high-quality, structured prompts using meta-prompting best practices: task decomposition, expert personas, iterative verification, and hallucination minimization.
Workflow
Bước 0: Detect Mode
| User cung cấp | Mode | Hành động |
|---|---|---|
| Mô tả yêu cầu (chưa có prompt) | Create | Chạy Phase 1 → 2 → 3 → 4 |
| Prompt có sẵn cần cải tiến | Refine | Chạy checklist Phase 4 trước → xác định điểm yếu → sửa theo kết quả |
Phase 1: Gather Requirements
Ask the user (one at a time, maximum 4 questions). Skip when answers are obvious from context.
- Type: Detect or ask — what type of prompt?
- System prompt: Persistent AI behavior across a conversation - Task prompt: One-shot, input → specific output - Conversational prompt: Multi-turn, requires memory/context handling - Role prompt: Character/persona, tone-focused
- Goal: "What is the primary goal or role of the system you want to create?"
- Output: "What specific outputs do you expect? (format, length, style)"
- Accuracy: "How should it handle uncertainty? (disclaim, ask for sources, or best-effort)"
Minimize friction — 1-2 questions usually suffice when context is clear.
Phase 2: Decompose (if needed)
| Tín hiệu | Decompose? |
|---|---|
| Prompt yêu cầu >= 3 expert domains khác nhau | Yes |
| Output cần cả phân tích + sáng tạo + validation | Yes |
| Single domain, straightforward input → output | No — skip to Phase 3 |
Khi decompose, assign expert personas — mỗi expert nhận instructions riêng, không shared memory:
- Expert Writer — copywriting, narrative, tone
- Expert Analyst — data, logic, verification
- Expert Python — code generation, computation
- Expert [Domain] — specialized knowledge
"Fresh eyes" — không assign cùng expert cho cả create AND validate.
Phase 3: Generate the Prompt
Consolidate vào 1 prompt cohesive. Dùng [prompt-template](references/prompt-template.md) — chỉ include sections phù hợp, omit phần không relevant.
Lưu ý khi generate:
- Match prompt structure theo prompt type (system vs task vs conversational vs role)
- Ưu tiên "must NOT" over "should" — negative framing hiệu quả hơn vì giảm ambiguity
Phase 4: Verify and Deliver
Self-review checklist trước khi deliver:
- No vague virtue words ("good", "helpful", "detailed") without specific definition
- No contradictory instructions (e.g., "be concise" + "include all details")
- No implicit assumptions — all context is explicitly stated
- Negative constraints are included (what NOT to do)
- Prompt type matches the use case
- If complex: suggest splitting into multi-prompt pipeline instead of mega-prompt
Then:
- Present the final prompt, organized and easy to follow
- Note expert reviews if used
- Offer to iterate if user wants adjustments
Anti-patterns to Avoid
When reviewing or generating prompts, flag these issues:
- Vague virtues: "Write a good summary" → "Write a 3-sentence summary focusing on key findings and action items"
- Overloaded prompt: Too many tasks in one prompt → suggest splitting
- Sycophancy traps: "Always agree" or "Never say no" → remove
- Missing scope: "Write marketing copy" without audience, channel, tone → clarify
- Prompt stuffing: Repeating the same instruction in multiple sections → deduplicate