Anthropic
The one family with a documented preference for XML structure, and the one that responds most to being told why an instruction matters.
How to prompt it
Wrap each kind of content in its own tag: <context>, <instructions>, <examples>, <input>, <output_format>.
Put role and context before the task. Anthropic documents context-first ordering as the reliable arrangement.
Explain why a rule exists. "Never use ellipses" works less well than "this is read aloud by a text-to-speech engine, which cannot pronounce them".
Include three to five examples, wrapped in <example> tags, varied enough that the model does not latch onto an accidental pattern.
Ask for exceptional effort explicitly if you want it — the model will not infer it from an enthusiastic prompt.
Say how long the answer should be. Left alone, it errs long.
Strengths
- Parses XML-tagged prompts unusually reliably, which keeps instructions, examples and data from bleeding into each other.
- Generalises from explained instructions: give the reason and it applies the rule to cases you did not list.
- Long context across the current lineup, with strong long-document handling.
- An effort setting that lets you dial thinking depth without rewriting the prompt.
Weaknesses
- Verbose by default. Anthropic’s own docs tell you to steer output length explicitly if you want it shorter.
- Vague prompts get conservative, hedged answers; "go beyond the basics" has to be asked for.
- The XML habit does not transfer cleanly — the same prompt sent to a chat-first assistant can read as clutter.
Models
Claude Fable 5
Next-generation intelligence for long-running agents.
- Context
- 1M
- Max output
- 128K
Good for: The hardest reasoning and long-horizon agent work.
Watch out: Slowest and most expensive of the lineup; thinking is always on.
Claude Opus 5
For complex agentic coding and enterprise work.
- Context
- 1M
- Max output
- 128K
Good for: Coding, analysis, and multi-step professional work at moderate latency.
Watch out: Give it a length budget for written deliverables or it will over-deliver.
Claude Sonnet 5
The best combination of speed and intelligence.
- Context
- 1M
- Max output
- 128K
Good for: Day-to-day drafting, summarising and interactive use.
Watch out: Follows instructions literally — say what you mean rather than hinting.
Claude Haiku 4.5
The fastest model with near-frontier intelligence.
- Context
- 200K
- Max output
- 64K
Good for: High-volume classification, extraction and quick turns.
Watch out: Smaller context window than the rest, and no effort setting.
Sources and grounding
- 1Prompting best practicesAnthropic (Claude Platform docs)The reference this page’s rules are drawn from, including XML structuring, roles, examples and context-first ordering.
- 2Models overviewAnthropic (Claude Platform docs)Live lineup, context windows, output limits and pricing.