Claude — System Prompts, Long Context, XML Tags

What Claude does well

Long context (1M tokens on the latest Opus tier).
Extended thinking (explicit budget, visible-to-operator thinking blocks).
XML-tagged prompts — the model is heavily trained to respect tag boundaries.
Tool use with parallel calls and rich error semantics.
Citation feature for grounded answers with span-level traceability.

Quirks to know

System prompt is a top-level parameter, not a message — don't mix.
Last message must be a user message (or the model has nothing to respond to). Prefill is enabled by appending an assistant message that the model continues.
Token counting is server-side via the messages.count_tokens endpoint — there isn't a perfect client-side tokenizer.
JSON mode is via tools (tool_choice with a single output tool whose input_schema = your output schema).

Style preferences

Claude responds well to explicit structure: numbered lists, XML tags, named sections. Vague tone instructions ("be helpful") matter less than format anchors and field semantics.

Code

Claude — full-feature call·python

import anthropic

client = anthropic.Anthropic()
resp = client.messages.create(
    model="claude-opus-4-7",
    system=[
        {"type": "text", "text": LARGE_REF, "cache_control": {"type": "ephemeral"}},
        {"type": "text", "text": "You are a contract analyst. Cite sources from <docs>."}
    ],
    thinking={"type": "enabled", "budget_tokens": 6_000},
    tools=[{"name": "emit_findings", "input_schema": SCHEMA}],
    tool_choice={"type": "tool", "name": "emit_findings"},
    max_tokens=2048,
    messages=[{"role": "user", "content": question}]
)

Claude — System Prompts, Long Context, XML Tags

What Claude does well

Quirks to know

Style preferences

Code

External links

Exercise

Progress

Comments 0