The official advice, minus the fluff.

Clarity, context and structure for Claude models, plus agentic and long-horizon prompting.

• Be clear and direct: treat Claude like a brilliant new employee who lacks your context.
• Add the reason behind instructions; Claude generalises from explanations.
• Use XML tags to separate instructions, documents and examples.
• Place long documents at the top of the prompt and the query at the end.
• Current models follow instructions precisely, so dial down aggressive emphasis that older models needed.
• For long agent runs, have the model keep state in files and verify its work with tests.

Best for: Anyone writing system prompts or agent briefs, on any model.

Instruction hierarchy, reasoning models, structured outputs and agentic workflows for GPT models.

• Put durable instructions in developer or system messages; keep user messages for the task.
• Reasoning models work best with high-level goals; avoid telling them how to think.
• Use Structured Outputs with a JSON schema when a program reads the result.
• Markdown headers and XML tags help the model parse long prompts.
• For agents, tell the model to keep going until the task is fully resolved and to use tools rather than guess.
• Pin model versions and evaluate prompt changes against a test set.

Best for: Developers building on the OpenAI API and agent builders.

Clear instructions, few-shot examples, context and multimodal prompting for Gemini.

• Always include few-shot examples when you can; prompts without them are less effective.
• Show positive patterns to follow rather than anti-patterns to avoid.
• Use prefixes (Input:, Output:) to signal structure.
• Break complex tasks into components and chain them.
• For images and video, put the media before the text instruction.
• Experiment with temperature and other parameters as part of prompt design.

Best for: Multimodal prompts and teams building on Gemini.

How prompting relates to model internals, and when to move from prompts to tuning.

• Prompts steer a fixed model; parameter-efficient tuning changes behaviour more durably.
• Few-shot prompting trades context length for task performance.
• Chain-of-thought helps on multi-step arithmetic and logic.
• When prompts get long and brittle, consider prompt tuning or fine-tuning instead.
• Measure on a held-out set; prompt gains often do not transfer across models.

Best for: ML engineers deciding between prompting, RAG and fine-tuning.

Practical prompting for educators and professionals new to generative AI.

• Provide context: who you are, who the output is for, and why.
• Be specific about format, length and tone.
• Build on the conversation; refine instead of starting over.
• Ask the model to take on a perspective relevant to the task.
• Always verify facts and citations independently.

Best for: Beginners and educators.

A beginner-friendly checklist for clear, safe prompting in academic and work settings.

• Be specific about the task and desired output.
• Give the AI a role and an audience.
• Use do and don't lists to set boundaries.
• Ask for multiple options and choose.
• Do not paste confidential data into tools your institution has not approved.

Best for: Students, staff and anyone in a regulated environment.

Getting fast, precise results from Grok, especially for agentic coding.

• Provide the relevant context explicitly: file paths, dependencies and code.
• Set explicit goals and requirements; vague prompts waste iterations.
• Use fast models for quick iteration and refine with follow-ups.
• Lean on native tool calling rather than describing tools in prose.
• Detailed system prompts measurably help in agentic setups.

Best for: Developers using Grok for coding and agent tasks.

Prompting open-weights Llama models, including chat templates and limitations.

• Use the model's official chat template exactly; formatting errors degrade quality.
• Explicit instructions about style and length work well.
• Few-shot and chain-of-thought help smaller models more than large ones.
• Restrict extraneous tokens by specifying output format.
• Expect to tune prompts per model size.

Best for: Teams self-hosting open-weights models.