Your best proposals take 12 hours to write. This one took 3 minutes.

# The decomposition agent reads the wizard inputs and the agent registry,
# then selects which specialists to run. Structured JSON output only:
# no prose, no reasoning outside the schema.
_SYSTEM_PROMPT = (
    "You are the decomposition agent for an intelligent document generation system.\n\n"
    "Your sole responsibility is to review the user's wizard inputs and the available "
    "specialist agent registry, then select the most appropriate agents to generate "
    "the requested document.\n\n"
    "Selection rules:\n"
    "1. Only select agents whose 'when_to_use' description matches the document type.\n"
    "2. Do not select duplicate agents.\n"
    "3. Always include at least one agent.\n"
    "4. Prefer quality over quantity: only select agents that add genuine value.\n"
    "5. Return your response as valid JSON matching the provided schema exactly.\n\n"
    "## Gap Detection\n\n"
    "After selecting agents, check if there is a significant capability gap: "
    "a need that no existing agent covers. Only flag genuine gaps, not minor "
    "nuances within an existing agent's scope.\n\n"
    "Place all reasoning inside each agent's 'reason' field. "
    "No explanation outside the JSON structure."
)

async def execute(self, input: DecompositionInput) -> DecompositionOutput:
    response = await self._llm.generate(
        LLMRequest(
            model="claude-haiku",       # alias for Claude Haiku 4.5: fast, cheap, structured selection
            system_prompt=_SYSTEM_PROMPT,
            user_message=self._build_user_message(input),
            output_schema=_DECOMPOSITION_OUTPUT_SCHEMA,
            max_tokens=2048,
            temperature=0.3,            # low temp: consistent routing decisions
        )
    )
    return self._parse_output(response)