CRISPE 🟡
Need: AI should act as an expert and provide a comprehensive, personalized response.
Explanation: Capacity (role) → Role (task) → Insight (context) → Statement (instruction) → Personality (style) → Experiment (variant). Defines the full framework for a task.
Prompt format:
“You are [role]. Your task is [goal]. Know that [context]. Write [what]. In the style of [style]. Also suggest an alternative.”
QA usage example
“You are a senior QA engineer. Your task is to review a test plan. Know that the project uses Java 17 + Cucumber. Write a list of observations. Be specific and technical. Also suggest an alternative test breakdown.”
When to use CRISPE?
CRISPE works best when you need a full, expert-level response — document reviews, test architecture analysis, strategy preparation. It sets context, role, and style in a single prompt.
Chain of Thought (CoT) 🟢
Need: The task requires logical reasoning or multi-step analysis.
Explanation: The model “thinks out loud” — it generates intermediate reasoning steps before giving an answer. Reduces hallucinations and logical errors.
Prompt format:
“[Question/task]. Think step by step.”
QA usage example
“A user reports that tests pass locally but fail on CI. Analyze the possible causes step by step, starting from the simplest.”
When to use CoT?
Whenever the problem requires analysis — debugging, root cause analysis, comparing options. Four words — “think step by step” — can dramatically improve response quality.
In the next post: RAG and Few-Shot — how to provide AI with context and teach it your expected output format.