Task decomposition prompt builder
Large language models get less reliable the more you cram into one prompt. A single request that asks the model to research, analyse, draft, and format all at once produces muddled output and compounding errors. Decomposition fixes this: split the task into focused steps, run them in sequence, and feed each result into the next. This builder generates that chain from a plain description of your goal.
How it works
You describe the task and pick how many steps to split it into. The builder produces a numbered chain where every step states the overall goal, its own narrow job, and a clearly marked handoff slot for the previous step’s output. You choose the handoff style — full, which carries the entire prior result forward, or summary, which asks you to condense it first to keep long chains cheap. Running the steps in order gives each model call a clean, single-purpose context instead of one overloaded prompt.
Why decomposition improves accuracy
The underlying reason decomposition works is that language models are essentially completing a pattern from their context window. When the context contains a single, clearly defined task, the model can focus all its probability mass on getting that one thing right. When the context contains five interleaved tasks, it must track multiple competing objectives simultaneously and errors in one thread bleed into others.
Research and anecdote from practitioners consistently finds that:
- Multi-step chains reduce “task loss” — where the model completes some sub-tasks and silently drops others.
- Reviewing output between steps lets you catch and correct an error before it propagates through all remaining steps.
- Smaller steps produce shorter, more checkable outputs, which are easier to verify than a single enormous document.
Designing a good chain
The most common mistake is making steps too coarse — steps that each contain multiple distinct jobs. A well-formed step has exactly one verb:
| Too coarse | Better |
|---|---|
| ”Research the topic and write a summary” | Step 1: Research. Step 2: Summarise. |
| ”Extract data and then analyse trends” | Step 1: Extract. Step 2: Analyse. |
| ”Draft and edit the copy” | Step 1: Draft. Step 2: Edit. |
Three to five steps covers the majority of complex tasks. More than seven steps adds handoff overhead and token cost without a proportional accuracy gain, and usually signals that the original task needs to be broken into separate tasks rather than a deeper chain.
Full handoff vs. summary handoff
- Full handoff: paste the entire previous step’s output into the next prompt. Best for short outputs or when later steps need every detail.
- Summary handoff: condense the prior output to the facts the next step actually needs, then paste that. Keeps context tight, reduces cost, and often improves focus on long chains.
Once a chain reliably produces good output, the same prompts drop into code with programmatic handoffs — replacing your clipboard with function calls that pass outputs between API requests.