AI Prompt Engineering

Role–Task–Format

The three essential components every effective prompt must specify.

Role definition

Sets expected expertise

Task instructions

Tell the model exactly what to do and how to approach the task.

Output format specification

Ensures answers are structured

Anchoring

Placing key constraints at the beginning and end to increase adherence.

Grounding

Forcing the model to rely on provided text/data instead of general knowledge.

Chain-of-thought (CoT)

Requesting explicit intermediate reasoning steps for complex tasks.

When to avoid CoT

When concise answers are needed or hallucination risk increases.

Few-shot prompting

Providing examples to teach a pattern or structure.

Least-to-most prompting

Breaking a complex task into smaller ordered subproblems.

Tree-of-thoughts

Exploring multiple reasoning branches and picking the best path.

Self-consistency

Sampling multiple reasoning paths and selecting the majority answer.

ReAct

Interleaving reasoning with actions or tool use based on observations.

Metacognitive prompting

Instructing the model to reflect on assumptions or missing information before answering.

Purpose of metacognition

Improves reasoning reliability and reduces subtle errors.

Self-critique prompting

Having the model evaluate its own output for flaws or gaps.

Self-refinement loop

A cycle of answer → critique → revise to improve output quality.

Alternative-consideration prompting

Asking the model what other interpretations or solutions it considered.

Confidence scoring

Requiring the model to report uncertainty to reduce fabricated answers.

Constraints

Rules that limit length

Output determinism

Using structure plus low temperature to ensure consistency.

Format enforcement

Requiring JSON

Dual placement of constraints

Putting rules at start and end to increase adherence.

Prompt compression

Expressing instructions in fewer tokens without losing structure.

Evaluation via examples

Testing prompts on example sets to ensure correctness and robustness.

Debugging prompts

Testing each requirement separately to isolate failures.

Asking for reasoning explanation

Reveals misunderstandings behind incorrect answers.

Hallucination trigger

Any task requiring explanation without grounding.

Hallucination prevention

Forcing model to cite provided text or source locations.

Constraint failure cause

Constraints placed only in the middle or conflicting instructions.

Fixing structural drift

Provide explicit templates and require exact replication.

Ambiguity resolution

Add minimal few-shot examples to clarify vague tasks.

Retrieval-augmented prompting

Providing documents or data that the model must use to answer accurately.

Prompt chaining

Solving a task as multiple sequential prompts that refine one another.

Constitutional principles

Rules guiding model behavior for accuracy

Agentic planning

Instructing the model to outline a step-by-step plan before acting.

Tool-aware prompting

Designing prompts assuming the model can call functions or APIs.

Ensemble prompting

Generating multiple answers and merging

Schema-based prompting

Using predefined templates or structures filled programmatically for consistent outputs.