Clear Instructions and Tool Boundaries for Low-Latency Tasks with Claude Haiku 4.5

One-sentence takeaway

Haiku 4.5's strengths are speed and cost efficiency; prompts should directly define scope, output shape, and the boundaries for tool execution. Using XML delimiters for inputs and examples can reduce misunderstandings in complex contexts for lightweight models.

Use cases

• Suitable tasks: Customer-service classification, summarization, rapid code drafting, sub-agents, batch extraction, and low-latency tool calls.

• Unsuitable tasks: Deep architectural reasoning, highly complex multi-file refactoring, and critical security decisions; route these to Sonnet/Opus and retain human or automated verification.

• Applicable model version: Claude Haiku 4.5, with the API name claude-haiku-4-5.

• Applicable clients, agents, or APIs: Claude API, Claude Code, Bedrock, and Vertex AI; the specific capabilities available depend on the platform.

• Recommended reasoning tier and parameters: Haiku 4.5 is a fast model. The official model table lists extended thinking as available and adaptive thinking as unavailable; start with standard requests and explicit output constraints, then configure extended thinking for complex tasks.

Ready-to-use content

<role>
You are a fast, careful execution assistant. Handle only work within the scope of the request.
</role>

<task>
Goal: <one-sentence-goal>
</task>

<input>
<item id="1"><content>...</content><source>...</source></item>
</input>

<rules>
1. Answer only from <input> and explicitly provided tool results; write "not found" for missing fields.
2. Complete the task first, then return the result in the format below; do not expand into unrequested functionality.
3. Call the search tool only when external or user-specific facts are genuinely needed.
4. Call the write tool only when the write conditions are met; read back and verify after writing.
5. If the task exceeds your reliable scope, explain why and recommend upgrading the model instead of making up an answer.
</rules>

<output_format>
{
  "result": "...",
  "evidence": ["source-or-item-id"],
  "uncertainties": ["..."],
  "next_action": "..."
}
</output_format>

Testing/workflow steps

1. Prepare normal cases, missing-field cases, and conflicting-constraint cases for the same batch of inputs.

2. Fix max_tokens, tools, and sampling parameters, then compare a natural-language prompt with an XML/JSON-constrained prompt.

3. Record JSON compliance rate, factual support, average latency, output tokens, number of tool calls, and manual review rate.

4. Upgrade failed or complex tasks to Sonnet according to the rules instead of making Haiku handle unsuitable work through repeatedly appended prompts.

Original evidence and data

• Anthropic's general guide recommends specifying the required output format and constraints; use numbered steps when order and completeness matter.

• The official recommendation is to separate long documents with XML structures such as <document>, <document_content>, and <source>, and to place the long document before the query.

• The official tool-use guide says that if you want Claude to take an action, explicitly write “use the tool to complete this” rather than only “give me some advice”; however, overly forceful MUST/CRITICAL language may lead to over-triggering.

• The official model page describes Haiku 4.5 as the fastest model with near-frontier intelligence, a 200k context window, and a 64k maximum output; extended thinking is available.

Applicability boundaries

• XML/JSON structures improve readability but do not automatically guarantee valid output; the server must perform parsing and schema validation.

• Fast models can still fail because of conflicting instructions, long contexts, or complex dependencies; an undisclosed success rate must not be fabricated.

• Tool triggering is behavioral guidance, not access control; tools should have an allowlist, timeout, sandbox, and post-write verification.

• When fresh information is needed, a search tool must be provided or the knowledge boundary must be stated clearly; Haiku must not be left to guess.

Source excerpt or observation (for compliant short quotation only)

Anthropic's general principle is “Be clear and direct” (a compliant short quotation); this is especially important for low-latency routing such as Haiku.