Rules Overview

PromptLint ships 21 rules covering security, cost, structure, and quality. Every rule can be individually enabled, disabled, or have its severity overridden.

All Rules

Rule	Severity	Auto-Fix	Category
cost	INFO	—	💰 Cost
cost-limit	CRITICAL	—	💰 Cost
prompt-injection	CRITICAL	✅	🔒 Security
jailbreak-pattern	CRITICAL	—	🔒 Security
secret-in-prompt	CRITICAL	—	🔒 Security
pii-in-prompt	CRITICAL	—	🔒 Security
context-injection-boundary	CRITICAL	—	🔒 Security
structure-sections	WARN	✅	🏗️ Structure
role-clarity	WARN	—	🏗️ Structure
output-format-missing	WARN	—	🏗️ Structure
output-length-missing	INFO	—	🏗️ Structure
hallucination-risk	WARN	—	🏗️ Structure
clarity-vague-terms	WARN	—	✨ Quality
specificity-examples	INFO	—	🎯 Specificity
specificity-constraints	INFO	—	🎯 Specificity
politeness-bloat	WARN	✅	📝 Verbosity
verbosity-redundancy	INFO	✅	📝 Verbosity
verbosity-sentence-length	INFO	—	📝 Verbosity
actionability-weak-verbs	INFO	—	💪 Actionability
consistency-terminology	INFO	—	🔄 Consistency
completeness-edge-cases	INFO	—	✅ Completeness

By Category

🔒 Security

The five security rules fire at CRITICAL severity by default. They should never be disabled in production pipelines.

Rule	What it catches
prompt-injection	Classic injection phrases like "ignore previous instructions"
jailbreak-pattern	Jailbreak attempts like "DAN mode", roleplay overrides, persona hijacking
secret-in-prompt	API keys, passwords, connection strings, tokens
pii-in-prompt	Email addresses, phone numbers, SSNs, credit card numbers
context-injection-boundary	User-injected content without a trust boundary marker

💰 Cost & Tokens

Rule	What it catches
cost	Token count and projected cost (INFO, always informational)
cost-limit	Prompt exceeds configured token_limit

🏗️ Structure

Rule	What it catches
structure-sections	No task / context / output sections detected
role-clarity	No role or persona defined for the model
output-format-missing	No output format specification
output-length-missing	No output length constraint (word count, sentence limit, etc.)
hallucination-risk	Patterns that increase hallucination likelihood

✨ Quality

Rule	What it catches
clarity-vague-terms	Vague words: "good", "efficient", "things", "various", "proper"
specificity-examples	No examples / few-shot demonstrations provided
specificity-constraints	No constraints on what the model should NOT do
politeness-bloat	Unnecessary politeness tokens ("please", "kindly", "thank you")
verbosity-redundancy	Redundant phrases ("in order to", "as well as")
verbosity-sentence-length	Sentences over 40 words
actionability-weak-verbs	Passive voice and weak verbs ("be done", "be written")
consistency-terminology	Same concept referred to by multiple names
completeness-edge-cases	No edge case handling defined

Severity Levels

Severity	Default exit code	Meaning
CRITICAL	2	Security or hard constraint violation — must fix
WARN	1	Structural or quality issue — should fix
INFO	0	Optimization suggestion — consider fixing

Override exit code threshold with --fail-level:

promptlint --file prompt.txt --fail-level warn  # exit 1 on any WARN or above
promptlint --file prompt.txt --fail-level info  # exit 1 on any finding

Disabling Rules

# .promptlintrc
rules:
  completeness_edge_cases: false
  actionability_weak_verbs: false