PM-301f · Module 1
When to Decompose
4 min read
The single-prompt limit is not primarily a token limit — it is a coherence limit. A prompt that asks a model to do too many distinct things in a single call produces outputs that are coherent on none of them. The model cannot maintain complete fidelity to ten simultaneous requirements across a long generation. It satisfies some and drifts on others.
The diagnostic question is not "how long is this task?" It is "how many distinct things does this task require?" A single task that requires synthesis of multiple document sources, construction of multiple distinct output formats, and application of multiple independent rule sets is a decomposition candidate — not because it is long, but because its requirements are heterogeneous.
- Signal 1: Ambiguous Scope If you cannot state the task's output in a single sentence without qualifications, the scope is ambiguous. Ambiguous scope in a single prompt produces ambiguous output. Decompose: separate the scope clarification from the execution, or decompose the execution into scoped sub-tasks.
- Signal 2: Multiple Output Types A task that requires both a JSON summary and a narrative explanation and a numbered action list is three tasks. Each output type has different format requirements. Combining them in a single prompt produces inconsistent compliance on all three. Separate them.
- Signal 3: Long Reasoning Chains Tasks that require more than 5-7 reasoning steps before reaching a conclusion are prone to error accumulation. Each reasoning step introduces potential for error. Decompose into checkpointed sub-tasks where each intermediate conclusion can be validated before the next step depends on it.
- Signal 4: Contradictory Constraints When a task has constraints that pull in opposite directions (maximize detail AND minimize length; be comprehensive AND be concise), the model negotiates a compromise rather than satisfying either. Decompose: separate the comprehensive draft from the condensed summary. Both can be done well; neither can be done simultaneously.
Do This
- Decompose when the task has multiple distinct output types
- Decompose when the reasoning chain is long enough that intermediate errors compound
- Decompose when constraints are heterogeneous across sections of the task
- Keep the task in a single prompt when the output is homogeneous and the reasoning is linear
Avoid This
- Decompose every task by default — decomposition has overhead; apply it when it solves a problem
- Decompose based on task length alone — a long single-output task may not need decomposition
- Decompose without mapping dependencies first — incorrect dependency order propagates errors
- Over-decompose into trivial single-sentence sub-tasks that each make one API call