When one prompt should not do everything

From one prompt to a workflow or agent

Use multiple stages when an intermediate result must be inspected, approved, logged or used to choose the next action.

Bound

Define the goal, permissions, risks and completion criteria.

Retrieve

Load only the sources and tools this stage needs.

Produce

Create an explicit intermediate deliverable rather than hidden state.

Check

Test against the requirements and real tool state before the next action.

Use the simplest mechanism that works

An agent is useful when the next step depends on a new finding or tool result. Deterministic code is better for fixed transformations, sorting, validation and calculation.

One prompt

Best for a bounded task with one body of material and one deliverable.

Fixed chain

Best when the stages are known: retrieve → analyse → write → check.

Agent

Best when the model must choose a tool or route based on what it discovers.

Write a contract for every stage

Stage: [name]
Goal: [one visible result]
Input: [exact material and earlier decisions]
Allowed tools: [smallest necessary set]
Not allowed: [side effects, data sources or files outside scope]

Deliver:
- result in [format]
- sources or tool evidence
- open questions and errors

Stop when: [measurable condition].
Ask for approval before: [external, destructive, costly or sensitive action].

Keep the handoff small

Pass forward decisions, evidence, open questions and required identifiers. Do not copy the entire reasoning trace or every raw tool result when a controlled summary is sufficient.

Evaluate the final product and the transitions

An intermediate stage can be correct while the final answer loses a source or requirement. Test the tool route, the intermediate deliverable and the visible final response.