A fast-reference guide to the AI concepts we use across prompts, agents, evaluations, and AI-assisted testing workflows.
Start with the language we use every day in AI courses and tools.
LLM: large language model Prompt: the input given to the model Context: supporting information sent with the prompt Completion: the model's output Token: a chunk of text the model processes
Strong prompts are clear about role, goal, context, constraints, and output format.
Role: You are a senior QA engineer Task: Generate regression test scenarios Context: Login module with OTP flow Constraints: Include positive, negative, boundary cases Output: Markdown table
Lower temperature is better for deterministic work like test cases, reviews, and summaries.
Low temperature: - more stable - less creative - good for structured outputs High temperature: - more varied - more exploratory - good for brainstorming
Models can sound confident and still be wrong. This is why grounding and evaluation matter.
Reduce hallucinations by: - adding trusted context - asking for citations - using structured outputs - validating responses with checks - keeping prompts specific
These patterns show up repeatedly in QA and SDET usage.
- Summarize a bug report - Generate test cases from requirements - Review Playwright code for flakiness - Explain API failures from logs - Convert notes into interview Q&A
AI outputs need testing too. Do not treat “sounds good” as “is good.”
Check for: - relevance - correctness - completeness - faithfulness to source context - safety / policy compliance - consistency across runs
Good QA prompts narrow the task and ask for coverage, reasoning, and output structure.
You are a senior SDET. Generate test scenarios for password reset. Cover positive, negative, boundary, security, and usability cases. Return as a markdown table with priority and preconditions.
These are the mistakes teams hit when they move too fast with AI tools.
- vague prompts - no validation - leaking secrets - overtrusting generated code - ignoring failure analysis - mixing stale and unverified context
AI should speed up work, not lower the quality bar.
- review generated output - keep prompt libraries - measure real usefulness - save strong examples - add human checkpoints - test AI workflows like any other feature