prompt-engeneering（即时工程）

Prompt Engineering

Universal techniques for crafting effective prompts across any LLM.

Core Principles

1. Structure with XML Tags

Use XML tags to create clear, parseable prompts:

<context>Background information here</context>
<instructions>
1. First step
2. Second step
</instructions>
<examples>Sample inputs/outputs</examples>
<output_format>Expected structure</output_format>

Benefits:

• Clarity: Separates context, instructions, and examples
• Accuracy: Prevents model from mixing up sections
• Flexibility: Easy to modify individual parts
• Parseability: Enables structured output extraction

Best practices:

• Use consistent tag names throughout (<instructions>, not sometimes <steps>)
• Reference tags explicitly: "Using the data in <context> tags..."
• Nest tags for hierarchy: <examples><example id="1">...</example></examples>
• Combine with other techniques: <thinking> for chain-of-thought, <answer> for final output

2. Control Output Shape

Specify explicit constraints on length, format, and structure:

<output_spec>
- Default: 3-6 sentences or ≤5 bullets
- Simple yes/no questions: ≤2 sentences
- Complex multi-step tasks:
  - 1 short overview paragraph
  - ≤5 bullets: What changed, Where, Risks, Next steps, Open questions
- Use Markdown with headers, bullets, tables when helpful
- Avoid long narrative paragraphs; prefer compact structure
</output_spec>

3. Prevent Scope Drift

Explicitly constrain what the model should NOT do:

<constraints>
- Implement EXACTLY and ONLY what is requested
- No extra features, components, or embellishments
- If ambiguous, choose the simplest valid interpretation
- Do NOT invent values, make assumptions, or add unrequested elements
</constraints>

4. Handle Ambiguity Explicitly

Prevent hallucinations and overconfidence:

<uncertainty_handling>
- If the question is ambiguous:
  - Ask 1-3 precise clarifying questions, OR
  - Present 2-3 plausible interpretations with labeled assumptions
- When facts may have changed: answer in general terms, state uncertainty
- Never fabricate exact figures or references when uncertain
- Prefer "Based on the provided context..." over absolute claims
</uncertainty_handling>

5. Long-Context Grounding

For inputs >10k tokens, add re-grounding instructions:

<long_context_handling>
- First, produce a short internal outline of key sections relevant to the request
- Re-state user constraints explicitly before answering
- Anchor claims to sections ("In the 'Data Retention' section...")
- Quote or paraphrase fine details (dates, thresholds, clauses)
</long_context_handling>

Agentic Prompts

Tool Usage Rules

<tool_usage>
- Prefer tools over internal knowledge for:
  - Fresh or user-specific data (tickets, orders, configs)
  - Specific IDs, URLs, or document references
- Parallelize independent reads when possible
- After write operations, restate: what changed, where, any validation performed
</tool_usage>

User Updates

<user_updates>
- Send brief updates (1-2 sentences) only when:
  - Starting a new major phase
  - Discovering something that changes the plan
- Avoid narrating routine operations
- Each update must include a concrete outcome ("Found X", "Updated Y")
- Do not expand scope beyond what was asked
</user_updates>

Self-Check for High-Risk Outputs

<self_check>
Before finalizing answers in sensitive contexts (legal, financial, safety):
- Re-scan for unstated assumptions
- Check for ungrounded numbers or claims
- Soften overly strong language ("always", "guaranteed")
- Explicitly state assumptions
</self_check>

Structured Extraction

For data extraction tasks, always provide a schema:

<extraction_spec>
Extract data into this exact schema (no extra fields):
{
  "field_name": "string",
  "optional_field": "string | null",
  "numeric_field": "number | null"
}
- If a field is not present in source, set to null (don't guess)
- Re-scan source for missed fields before returning
</extraction_spec>

Web Research Prompts

<research_guidelines>
- Browse the web for: time-sensitive topics, recommendations, navigational queries, ambiguous terms
- Include citations after paragraphs with web-derived claims
- Use multiple sources for key claims; prioritize primary sources
- Research until additional searching won't materially change the answer
- Structure output with Markdown: headers, bullets, tables for comparisons
</research_guidelines>

Example: Before/After

Without structure:

You're a financial analyst. Generate a Q2 report for investors. Include Revenue, Margins, Cash Flow. Use this data: {{DATA}}. Make it professional and concise.

With structure:

You're a financial analyst at AcmeCorp generating a Q2 report for investors.

<context>
AcmeCorp is a B2B SaaS company. Investors value transparency and actionable insights.
</context>

<data>
{{DATA}}
</data>

<instructions>
1. Include sections: Revenue Growth, Profit Margins, Cash Flow
2. Highlight strengths and areas for improvement
3. Use concise, professional tone
</instructions>

<output_format>
- Use bullet points with metrics and YoY changes
- Include "Action:" items for areas needing improvement
- End with 2-3 bullet Outlook section
</output_format>

Prompt Migration Checklist

When adapting prompts across models or versions:

1. Switch model, keep prompt identical — isolate the variable
2. Pin reasoning/thinking depth to match prior model's profile
3. Run evals — if results are good, ship
4. If regressions, tune prompt — adjust verbosity/format/scope constraints
5. Re-eval after each small change — one change at a time

Quick Reference

Prompting Techniques Catalog

Comprehensive catalog of prompting techniques. Full details, examples, and academic references in references/prompting-techniques.md.

Prompting Fundamentals

LLM settings, prompt elements, formatting, and practical examples — see references/prompting-introduction.md. Covers:

• LLM Settings — temperature, top-p, max length, stop sequences, frequency/presence penalties
• Prompt Elements — instruction, context, input data, output indicator
• Design Tips — start simple, be specific, avoid impreciseness, say what TO do (not what NOT to do)
• Task Examples — summarization, extraction, QA, classification, conversation, code generation, reasoning

Risks & Misuses

Adversarial attacks, factuality issues, and bias mitigation — see references/prompting-risks.md. Covers:

• Adversarial Prompting — prompt injection, prompt leaking, jailbreaking (DAN, Waluigi Effect), defense tactics
• Factuality — ground truth grounding, calibrated confidence, admit-ignorance patterns
• Biases — exemplar distribution skew, exemplar ordering effects, balanced few-shot design

Prompt Audit / Review

When asked to audit, review, or improve a prompt, follow this workflow. Full checklist with per-check references: prompt-audit-checklist.md.

Workflow

1. Read the prompt fully — identify its purpose, target model, and deployment context (interactive chat, agentic system, batch pipeline, RAG-augmented)
2. Walk 8 dimensions — check each, note issues with severity (Critical / Warning / Suggestion):

1. Produce a report — issues table (dimension, check, severity, issue, fix) + rewritten prompt or targeted fix suggestions. Use the report template from the checklist reference.
2. For each issue, cite the relevant reference file so the user can dive deeper.

Quick Decision: Which Dimensions to Prioritize

• User-facing chatbot → prioritize Safety (#3), Hallucination (#4), Clarity (#1)
• Agentic system with tools → prioritize Safety (#3), Context (#5), Maintainability (#6)
• Batch/pipeline → prioritize Structure (#2), Evaluation (#8), Maintainability (#6)
• RAG-augmented → prioritize Context (#5), Safety (#3), Hallucination (#4)

Common Mistakes & Anti-Patterns

Three complementary layers — use the one matching your need:

Deep-dives by category — root causes, mechanisms, prevention checklists (from "The Architecture of Instruction", 2026):

Quick reference — 18-category taxonomy with MRPs, risk scores, case studies, action items: failure-taxonomy.md. Start here for an overview or to prioritize which categories to address first. Covers: control-plane vs data-plane model, heuristic risk scoring, real-world incidents (EchoLeak CVE-2025-32711, Mata v. Avianca, Samsung shadow AI).

How to measure & test — eval metrics, CI gating, red-teaming, tooling: evaluation-redteaming.md. Covers: TruthfulQA, FActScore, SelfCheckGPT, PromptBench, AILuminate, LLM-as-judge pitfalls, guardrail libraries, open research questions.

Model-Specific Guides

Each model family has unique parameters, gotchas, and patterns. Consult the reference for your target model:

• Claude Family — Opus/Sonnet 4.6: adaptive thinking (effort param), prefill deprecation (use Structured Outputs), tool overtriggering fix, prompt caching, citations, context engineering, agentic subagent patterns, vision, migration from 4.5
• GPT-5 Family — GPT-5/5.1/5.2: reasoning_effort param (defaults vary per version), verbosity API control, named tools (apply_patch), agentic eagerness templates, compaction API, instruction conflict sensitivity, migration paths
• Gemini 3 Family — Gemini 2.5/3/3.1: temperature MUST be 1.0, thinking_budget vs thinking_level, constraint placement (end of prompt), persona priority, function calling, structured output, multimodal, image generation
• GPT-5.2 Specifics — Compaction API code examples, web research agent prompt, full XML specification blocks