A structural element gives one concern in a prompt a defined place. Separating the task, background, variable inputs, boundaries, and expected response makes a prompt easier to reuse and diagnose than a single undifferentiated block of text. section 2.0
Six elements form the prompt’s working interface:
- Instructions define the task. section 2.1
- Constraints set verifiable limits on acceptable output. section 2.1
- Context supplies background needed to interpret the task. section 2.1
- Input parameters hold values replaced for each run, keeping variable data separate from stable instructions. section 2.1
- Output format specifies the response’s structure and style. section 2.1
- Delimiters mark embedded content so it is distinguishable from instructions. section 2.1
Start with instructions and an output format. Add other elements when the task needs them: constraints for checkable requirements, context for interpretation, parameters when values change between runs, and delimiters around embedded material. Static inputs need not become parameters. Fixed headings or fields can make a response predictable without requiring a strict machine-readable contract. section 2.8 section 2.6
Deployment decision brief
A release prompt places live error rate, latency, and canary signals inside <deployment_signals> tags and rollback rules inside <runbook_excerpt> tags. Its instruction requests a GO or NO-GO recommendation using only those inputs. Constraints require a word range, at least two concrete signals, and a rollback command for NO-GO. The output format fixes four sections—Recommendation, Rationale, Immediate Action, and Risks to Monitor—and their bullet counts. These labels make it possible to check the response against the supplied evidence and requested shape separately. section 2.8.2
For repeated work, retain the task logic and replace only parameter values, such as a feature request’s goal and acceptance criteria. Structural elements define sections and boundaries; linguistic choices tune wording, while prompt patterns add reusable behavior. When a result fails, inspect the relevant element rather than rewriting the whole prompt. section 5.2.1 section 2.0
