Advanced Prompt Engineering Techniques: Chain of Thought, Roles, and Structured Outputs

Once basic prompts work, hard tasks need reasoning scaffolds and machine-readable outputs. Advanced prompt engineering is less about secret phrases and more about decomposing work the way a careful analyst would.

Start with prompt fundamentals if you are new.

Chain of thought (CoT)

Ask the model to reason step by step before the final answer for math, multi-hop analysis, and planning. In production, prefer hidden reasoning patterns your provider supports, or ask for a short “reasoning summary” if you must log explanations for humans.

Think through the problem in numbered steps.
Then provide Final answer: … in one line.

Caution: long CoT increases tokens/cost and can leak sensitive intermediate thoughts into logs—redact carefully.

Self-critique and dual-pass prompts

1. Generate a draft
2. Critique against a checklist (accuracy, tone, policy)
3. Revise

Two-pass flows often beat one giant mega-prompt for quality writing, code review, and compliance checks.

Role specialization

Instead of “be helpful,” assign a narrow expert role with decision rights: “You are a staff SRE reviewing a runbook for production safety. Prefer conservative changes.” Roles bias style and risk tolerance.

Structured outputs (JSON / schemas)

For software systems, free text is a bug. Prefer:

• Provider JSON mode / schema-constrained decoding when available
• Explicit schema in the prompt with types and enums
• Validation in code (Zod, Pydantic, JSON Schema) with repair loops

Return ONLY valid JSON matching:
{
  "priority": "P0"|"P1"|"P2",
  "summary": string,
  "actions": string[]
}
No markdown fences.

Tool-aware and ReAct-style prompting

When the model can call tools (search, DB, tickets), instruct: reason → choose tool → observe → answer. Never invent tool results. Production agents need hard validation around tool arguments—see production agents.

Decomposition patterns that scale

• Map-reduce — chunk long docs, summarize parts, then merge
• Plan-then-execute — outline steps, then run each step with focused prompts
• Router — classify intent, then send to specialized prompt

When advanced techniques hurt

• Simple extraction tasks (overthinking adds errors)
• Latency-sensitive UX (prefer smaller models + short prompts)
• Unvalidated CoT used as legal or medical advice

Frequently asked questions

Should every prompt say “think step by step”?

No. Use CoT for multi-step reasoning; skip it for formatting and straightforward transforms.

Is few-shot always better?

Usually for classification and style matching. Poor examples teach the wrong pattern—curate carefully.