AgentSkillsCN

gpt-prompting

GPT-5.2（GPT-5级）模型的提示模式、可复用的系统提示块及迁移清单。适用于以下场景：(1) 为GPT-5.2生产代理撰写或优化提示；(2) 控制输出的冗长程度与格式；(3) 防止范围偏离（尤其是前端/UX及过度构建）；(4) 处理模糊性并减少幻觉；(5) 提升工具的接地性与结构化提取能力；(6) 将提示从GPT-5/5.1/4.1/4o迁移至GPT-5.2，同时保持稳定的推理效率。

SKILL.md
--- frontmatter
name: gpt-prompting
description: Prompt patterns, reusable system-prompt blocks, and a migration checklist for GPT-5.2 (GPT-5-class) models. Use when you need to (1) write/refine prompts for GPT-5.2 production agents, (2) clamp verbosity/output shape, (3) prevent scope drift (esp. frontend/UX and over-building), (4) handle ambiguity and reduce hallucinations, (5) improve tool grounding + structured extraction, or (6) migrate prompts from GPT-5/5.1/4.1/4o to GPT-5.2 with stable reasoning_effort.

GPT Prompting (GPT-5.2)

Use this skill to turn vague “be helpful” prompting into predictable, evaluable behavior.

If you need the source guides, see:

For the block library + examples, read: references/guide.md

Quick start (recommended flow)

  1. •State the job + constraints (what success is, what not to do).
  2. •Add a verbosity/output-shape clamp.
  3. •Add risk rails (ambiguity + hallucination guard).
  4. •If tools exist: add tool usage rules and a post-write change recap.
  5. •If extracting data: add an extraction schema with null-for-missing.

Drop-in prompt skeleton

Use as a starting point for system prompts / instruction blocks:

text
You are an expert assistant.

<output_verbosity_spec>
- Default: 3–6 sentences OR ≤5 bullets.
- Simple questions: ≤2 sentences.
- Complex tasks: 1 short overview paragraph, then ≤5 bullets tagged:
  What changed, Where, Risks, Next steps, Open questions.
- Avoid long narrative paragraphs; prefer compact bullets + short sections.
</output_verbosity_spec>

<uncertainty_and_ambiguity>
- If ambiguous/underspecified: ask up to 1–3 precise clarifying questions OR present 2–3 interpretations with labeled assumptions.
- Never fabricate exact figures, IDs, line numbers, or citations.
- Prefer “Based on the provided context…” over absolute claims when uncertain.
</uncertainty_and_ambiguity>

<tool_usage_rules>
- Prefer tools over memory whenever you need fresh/user-specific data.
- Parallelize independent reads when possible.
- After any write/update tool call, restate:
  What changed, Where, and validation performed.
</tool_usage_rules>

<scope_discipline>
- Implement EXACTLY and ONLY what the user asked.
- No extra features, no embellishments.
- If something is ambiguous, choose the simplest valid interpretation.
</scope_discipline>

Migration checklist (GPT-5/5.1/4.x → GPT-5.2)

  • •Make one change at a time: switch model first; keep prompts functionally identical.
  • •Pin reasoning_effort to match the old latency/depth profile (don’t rely on defaults).
  • •Run evals; only then tune (usually: verbosity clamp + scope discipline + ambiguity rails).

See references/guide.md for a compact mapping table.