Prompt Engineering Basics: 7 Principles for Daily Work

Prompt engineering is the work of designing a request so an AI model understands your intent accurately. The official guides from Anthropic, OpenAI and Google word things differently, but they stress the same principles: clarity, context, examples, structure and iterative revision. Getting those five habits down changes the quality of results in ChatGPT, Claude or Gemini alike.

What prompt engineering means

Prompt engineering means understanding how an AI model takes in text, then shaping your question to fit. Typical techniques include few-shot prompting (showing a few examples), chain-of-thought (having the model work through its reasoning) and assigning the model a specific role.

Practitioner articles tend not to treat it as a developer-only skill. Planners and marketers need the same ability to organize requirements and hand them to someone else, and writing a prompt is an extension of that ability.

Why it still matters as models get smarter

Anthropic’s documentation suggests thinking of Claude as a capable new hire who does not yet know your team’s habits and standards. The more specifically you say what you want, the better the result. Even a very good new employee brings back the wrong thing if they do not know the background of the task.

Google’s Gemini guidance points the same way. Do not expect the model to already know what it needs; put the instructions and information required to solve the problem inside the prompt. In Google’s example, writing only ‘my Wi-Fi isn’t working’ gets a weaker answer than describing the current state, such as ‘the indicator light on my Google Wi-Fi router is blinking slowly in yellow’.

Seven basic principles for everyday work

State the goal and the deliverable specifically. Instead of ‘polish this report’, write something like ‘make a summary for my manager covering key results, problems and next steps in that order, in a concise, objective tone’. The output changes accordingly.
Give context and reasons. Say who will read it, where it will be used and why a constraint exists. Anthropic’s documentation explains that when you include the reason, the model grasps the goal better and answers in a way that fits the situation.
Show examples. Anthropic recommends including 3 to 5 examples for Claude. Be careful in choosing them, though, because models also copy unwanted traits that happen to be in the examples.
Separate the structure. Keep instructions, reference material, examples and input data apart with delimiters such as ###, quotation marks or XML tags. Google also states that the order, labels and delimiters you use affect response quality.
Specify the output format. Name the table, bullet list, length, tone or JSON you want. If you dislike lists, saying ‘write it in paragraphs’ works better than ‘no lists’.
Break big tasks into steps. Rather than asking for a plan and a budget at once, set an order and request them one after another. Linking requests this way is called prompt chaining.
Review the result and revise. Anthropic itself admits that a first prompt rarely lands perfectly. Keep prompts that worked well as templates so the next task goes faster.

For a self-check, a question suggested by Anthropic is useful: if you handed only this prompt to a colleague who knows nothing about the project, could they do the job? If a sentence would confuse the colleague, it will confuse the AI too. For an example of applying these principles to real work, see how to record, summarize and review meeting minutes with AI.

Reasoning models are used differently

Item	Anthropic (Claude)	OpenAI (reasoning models)	Google (Gemini)
Core direction	Be clear, as if briefing a new hire, and give the reasons	Short and direct	Include all needed information, with structure
Structuring	XML tags recommended	Use delimiters	Delimiters; separate elements such as goal and instructions
Examples	3 to 5 recommended	Try without examples first	Use few-shot to adjust format and scope
Step-by-step thinking	Extended thinking feature or CoT	Not needed	2.5 and 3 series think automatically

OpenAI’s guide for reasoning models explains that the model reasons internally on its own, so there is no need to add phrases like ‘think step by step’. It also advises against including examples from the start: run the prompt without them first and add examples only if the result falls short. It is worth knowing that when information is missing, reasoning models often ask the user a question rather than guess. The Gemini 2.5 and 3 series likewise go through an internal thinking process automatically, so you generally do not need to make the answer spell out each reasoning step.

Lately the term ‘context engineering’ appears often. It covers designing not just the question but also reference material, conversation history and tools. Anthropic sees context engineering as a natural extension of prompt engineering. A July 2025 Gartner report quoted by an industry outlet projected that context engineering would be applied to 80% of AI tools by 2028, but the same outlet concludes it is not a concept that replaces prompt engineering.

Cautions and limits

Responses differ by model and version. In its GPT-5.5 guidance, OpenAI urges treating it as a family whose prompts need to be retuned rather than dropped into the slot of an older model. If you switch models, recheck your existing prompts first.
A good prompt does not guarantee facts. Prompts can reduce hallucinations but cannot eliminate them, so ask for sources and have a person do the final check.
External documents may contain hidden instructions. When you paste in a web page or file, keep prompt injection in mind, and check the service’s data policy before entering confidential or personal information.
Frameworks such as CO-STAR and RISEN are not official terms. They come from communities and training providers, so treat them as reference only.

Frequently asked questions

Is prompt engineering only for developers?

No. The moment you type a request into the ChatGPT, Claude or Gemini chat window, you are writing a prompt. Developers building APIs or agents simply handle system prompts and context design on top of that.

Does the phrase ‘think step by step’ still work?

It depends on the model. OpenAI’s documentation says reasoning models do not need it, and the Gemini 2.5 and 3 series also think automatically on the inside. For general models without a reasoning feature, chain-of-thought prompting is still used.

Do I have to assign a role like ‘You are a marketer with 10 years of experience’?

It often helps but is not required. Some point out that if the model already understands the task properly, there is no need to layer a role on top. Spelling out the goal, the audience and the output format has a bigger effect on results than the role does.

From the author’s desk

Dave KimWeb developer & IT ops

Prompt writing is close to what a product planner does when handing requirements to a team. When I plan websites and custom business systems as a freelancer, vague briefs are what cause the most rework, and I expect the same holds for prompts. I’d keep a small template file of prompts that worked, so you can reuse them without rewriting. Since the post says models behave differently by version, I’d also re-run your saved prompts on a few real tasks whenever you switch models. And before pasting in internal documents, check what your company’s data policy allows.