FAQ

Does PromptLint send my prompts anywhere?

No. All analysis is local. PromptLint never makes network requests during linting. Your prompt text stays on your machine.

The one exception: if you use the tiktoken-based token counter (pip install "promptlint-cli[tiktoken]"), tiktoken downloads its vocabulary files on first use — but that's a one-time download of model vocabulary data, not your prompt text.

---

Python vs. Node.js — which should I use?

Feature	Python CLI	Node.js CLI
Token counting	Exact (tiktoken)	Heuristic (chars ÷ 4, ±15%)
Auto-fix	✅	✅
SARIF output	✅	✅
Rule coverage	20 rules	20 rules
Speed	~100ms startup	~80ms startup

Use Python if you care about exact token counts and cost projections. Use Node.js if you're in a JavaScript/TypeScript environment. For CI/CD, both work equally well — use the GitHub Action which wraps the Node.js CLI.

---

Are patterns matched as regex or plain substrings?

Regex. All patterns — injection patterns, jailbreak patterns, and custom patterns you add in .promptlintrc — are compiled as case-insensitive regular expressions.

Simple phrases like ignore previous instructions work as substring regex (no anchors = match anywhere in the string). You can use full regex syntax:

rules:
  prompt_injection:
    patterns:
      - "ignore previous instructions"           # substring match
      - "admin (override|bypass):"               # alternation
      - "(?:you are|act as) (?:an? )?[A-Z]\\w+" # group + character class

Invalid patterns are skipped, not errors

If a pattern fails to compile, PromptLint logs a warning and continues — it does not crash. Test custom patterns with promptlint --explain prompt-injection or by linting a test file.

---

Which models does PromptLint support?

PromptLint is model-agnostic — it analyzes the prompt text regardless of which model you use. The model config setting only affects token counting (via tiktoken) and cost projection.

Supported by tiktoken (for exact counts): gpt-4o, gpt-4, gpt-3.5-turbo, and other OpenAI models. For non-OpenAI models (Claude, Gemini, Llama), tiktoken falls back to cl100k_base encoding which is close but not exact for those tokenizers.

---

How do I suppress a false positive?

Option 1 — Disable the rule entirely:

rules:
  actionability_weak_verbs: false

Option 2 — Demote to a level below your --fail-level:

rules:
  specificity_examples:
    enabled: true
    level: info   # won't fail a --fail-level warn pipeline

Option 3 — Narrow the rule scope. For example, reduce injection patterns to only the most critical ones:

rules:
  prompt_injection:
    patterns:
      - ignore previous instructions
      - admin override:
      # (omitting broader patterns that cause false positives)

---

Does PromptLint handle template variables like ?

Yes — template variables are preserved in all fixes and most rules. The context-injection-boundary rule specifically looks for template variables and fires when it finds one without a trust boundary marker nearby.

Other rules treat as a literal string. For example, pii-in-prompt won't flag — only actual email addresses like user@example.com.

---

Can I run PromptLint on multiple files at once?

Yes. Three approaches:

# Glob pattern (quotes required on most shells)
promptlint --file "prompts/**/*.txt"

# Shell loop (more control per file)
for f in prompts/*.txt; do
  promptlint --file "$f" --format json
done | jq -s '.'  # merge JSON results

# GitHub Actions (scans a directory)
- uses: AryaanSheth/promptlint@v1
  with:
    path: prompts/

When using a glob, the exit code reflects the worst finding across all files.

---

What does the score mean?

The score (0–100) is computed across four weighted categories:

Category	Weight	Penalty per CRITICAL	per WARN	per INFO
Security	40%	−30	−10	−3
Cost	20%	−30	−10	−3
Quality	25%	−30	−8	−3
Completeness	15%	−30	−10	−3

Each category is scored 0–100, then combined by weight. Scores are clamped to 0.

Grades: A (90+), B (75+), C (60+), D (45+), F (below 45)

---

Why does the specificity-examples rule fire on my prompt even though I have examples?

The rule checks for the word example (or e.g., such as, like, for instance) anywhere in the prompt. If your examples don't use any of those keywords, the rule still fires.

Fix: add a label before your examples:

Examples:
  Input: ...
  Output: ...

Or add the <examples> tag. The word "example" anywhere in the text suppresses the rule.

---

Can I use PromptLint with non-English prompts?

Partially. Security rules (injection detection, PII, secrets) work on any language because they use regex patterns, not NLP. Quality rules (vague terms, politeness, redundancy) are English-optimized and will have high false-positive rates on other languages — disable those rules for non-English prompts.

---

Can I use PromptLint in Docker?

Yes. See the Docker integration guide for a minimal Dockerfile and docker-compose setup.

---

Where do I report bugs or request features?

Open an issue on GitHub. Include the prompt text (sanitized if needed), your .promptlintrc, the PromptLint version (promptlint --version), and the unexpected output.