OpenAI, Anthropic and Google are converging on an unglamorous prompting playbook: state the job clearly, supply missing context, show the pattern you want, define the output and split complicated work. For users who have watched a chatbot nail one request and fumble the next, that agreement matters more than another list of “100 prompts that change everything.”
- Clarity is the baseline: specify the task, limits and format instead of hoping the model infers them.
- Examples are the strongest lever: Google says include them whenever possible; Anthropic suggests 3–5 varied cases; OpenAI recommends a diverse range.
- Structure and test: separate instructions, data and examples; put long documents first and the question last; evaluate prompts like code.
Clarity, context and the reason behind the request
All three labs put clear instructions first. Anthropic suggests handing your prompt to a colleague with little context: if they would be lost, so will the model. Google adds constraints (“summarise this in one sentence”) and response format (table, list, paragraph, keywords), and warns against unnecessary or overly persuasive language. OpenAI says GPT models do better with explicit instructions.
Anthropic adds a practical rule that improves results more than most tricks: say what you want, not what you do not want. Instead of “do not use lists,” ask for “prose paragraphs.” Anthropic’s example is that “never use ellipses” works less well than explaining that a speech synthesiser will read the answer aloud and cannot pronounce them. Google frames this as adding the information the model needs; OpenAI notes that context can also point the model at data it never saw or restrict it to chosen sources.
Explain why: a constraint with its reason attached lets the model handle cases the prompt never anticipated.
With long documents, Anthropic and Google put the material first and the question last; Anthropic says that can improve response quality by up to 30% on complex, multi-document inputs. Google suggests a transition such as “based on the information above…” to link document and query.
Examples, roles and prompt architecture
If you want a tone or format, show it. Google says include few-shot examples whenever possible because its testing found them the most reliable lever; clear examples can even replace some instructions. The risk is overfitting, so Anthropic recommends 3 to 5 relevant, varied, clearly delimited examples. OpenAI asks for a diverse range of inputs and outputs. Pasting two or three summaries you like beats three paragraphs explaining your taste.
A short role and clear structure help. OpenAI orders system messages as identity, instructions, examples and context, using Markdown headings and XML tags. Anthropic says even a one-sentence role changes results and recommends tags for each content type; Google advises one consistent separator. Over-emphatic prompts age badly: OpenAI says reasoning models do not need step-by-step instructions; Google warns old complex techniques can cause overthinking; Anthropic prefers normal phrasing over “CRITICAL: You MUST” and says pre-filling answers is no longer accepted by its newest Claude models.
Split, evaluate and move prompts into code
Google recommends one prompt per instruction, chaining outputs into inputs or splitting data and aggregating results. Anthropic says its models handle much multi-step reasoning internally, but chaining remains useful for inspecting intermediate work; its common pattern is draft, review against criteria, then final version. OpenAI’s general guide does not stress splitting, but treats prompts as code: output quality depends on prompt quality, so teams should test and evaluate changes. The operational signal: OpenAI will begin winding down API prompt-saving from June 2026, with a full shutdown on November 30, pushing durable prompts into version-controlled code.
The evidence also argues against piling on tricks. A 2025 Wharton School review found tips, threats and expert-role instructions did not consistently improve hard-problem accuracy, while step-by-step requests gave reasoning models little gain and used more tokens. Google DeepMind research found AI-optimised prompts beating human-written ones by up to 8% on maths word problems and 50% on difficult sets. A 2024 VMware paper reported automatic optimisation as the most effective approach.
The mature lesson is less mystical than the old hype. The best requests resemble a good professional brief: objective, audience, relevant material, an example of excellence, delivery format and a verification step. Models usually will not ask for missing details, so the burden falls on the person writing the brief. Anthropic now frames the wider discipline as context engineering—connecting models to the right data, tools and constraints—rather than collecting incantations. That is where the real advantage is moving, and why clear, testable prompts remain the interface between vague intention and useful work.