What Anthropic, Google, and OpenAI all agree on about prompts

Prompt advice online is mostly vibes. But each lab publishes official guidance for its own models: Anthropic's best practices, Google's prompting strategies, OpenAI's prompt engineering guide, and Meta's Llama prompting guide. When four labs that compete on everything agree on a rule, that rule is as close to ground truth as prompt engineering gets.

1. Be explicit, and phrase instructions positively. Every lab says some version of this. Anthropic frames it memorably: show your prompt to a colleague with minimal context, and "If they'd be confused, Claude will be too." Google's prompt engineering whitepaper distills it as instructions over constraints: say what to do, not what to avoid. Tell the model why a rule exists and it generalizes instead of rules-lawyering.

2. Show examples (few-shot). Google pushes hardest here; its Gemini docs state that prompts without few-shot examples are "likely to be less effective." Anthropic recommends 3-5 diverse, structured examples. The nuance: keep every example in the exact same format, or the stray format leaks into output.

3. Separate sections with delimiters. Instructions, context, and input data should never blur together. Anthropic prefers XML tags, OpenAI starts with Markdown headings, Google accepts both and asks only for consistency. The agreement is structural, not syntactic: label your sections.

4. Give the model a role. A one-line persona ("You are a senior engineer reviewing a pull request") focuses tone and judgment. All four guides include it; Google's consumer formula is literally persona, task, context, format.

5. Iterate empirically. None of the labs treats a prompt as finished after one draft. Google's Vertex docs have an entire page named "Prompt iteration strategies" and its whitepaper tells you to document your attempts. OpenAI's GPT-5 guide recommends metaprompting: asking the model which phrases to add or delete. Prompts are drafts that improve under testing — which is, not coincidentally, why we built version history and Loop into this product.

Delimiters   Anthropic: XML-first    OpenAI: Markdown-first   Google: either, be consistent
CoT          classic models: yes     reasoning models: DON'T prompt "think step by step"
Prefill      was Anthropic-only — returns an error on Claude 4.6+; migrate to structured outputs
Placement    Anthropic/Google: query at the END of long context   OpenAI: instructions at BOTH ends

The practical takeaway: write prompts on the universal five, and treat the divergences as per-model settings. A prompt built from clear positive instructions, labeled sections, a role, and examples travels across Claude, Gemini, and GPT with minor edits. A prompt built on model-specific tricks (prefill, magic phrases) breaks on the next model release.

Every prompt in our public library follows the universal five — constraint-scoped, section-labeled, role-first, with output formats and numeric judgments. Two examples you can copy or fork: pr-review and threat-model-first. Each one is versioned, so when a lab updates its guidance, the prompt's history shows exactly what changed and why.