← All Tools
System-Prompt Linter
Paste the system prompt that drives your assistant or agent. The linter flags weak refusals, prompt-injection footguns, missing output anchors, contradictions, and credentials that should never be in a prompt.
100% client-side. Your prompt stays in the browser. The patterns and heuristics run locally.
⚠ Heuristic, not exhaustive. Lint catches patterns; it cannot prove a prompt is jailbreak-resistant. Always combine with adversarial testing (red teaming) and an output filter / classifier.
—
What the linter looks for
Role & framing
- Clear opening role declaration ("You are an X that does Y.")
- Stated capabilities and out-of-scope topics
- Concrete output format anchor (Markdown, JSON, <tool> tags)
Refusal & safety
- Explicit refusal trigger conditions
- What to say when refusing (canonical refusal phrasing)
- No "always answer / never refuse / be honest about everything" patterns
Prompt injection resilience
- Untrusted user / RAG content delimited and labelled
- Instructions inside untrusted content treated as data, not commands
- Tool / function descriptions clearly bounded
Hygiene
- No real credentials / keys / PII in the prompt itself
- No contradictions ("never X" + "always X")
- No "ignore previous instructions" baked in for testing and forgotten