Prompt engineering checklist
A pre-flight checklist before shipping a new agent prompt to production.
Before shipping a prompt
- Output format is explicitly specified (schema, not prose description)
- Hard constraints are stated immediately before the task, not buried earlier
- At least 3 edge cases have been manually tested
- The prompt is version-controlled with a changelog entry
- There’s a fallback behavior defined for malformed output
- Token budget is accounted for (input + expected output)
Common failure patterns
| Symptom | Likely cause |
|---|---|
| Inconsistent output structure | Format described in prose, not enforced |
| Ignores late instructions | Constraints buried too early in a long prompt |
| Overly verbose reviewer feedback | No concrete rubric given |
| Silent truncation | Output budget not reserved separately from input |