promptlint-cli 1.3.0

A static analyzer for LLM prompts — catches cost waste, quality issues, and security risks.

PromptLint

Static analysis for LLM prompts. Think ESLint, but for the text you send to GPT-4 / Claude / Gemini.

Catches token waste, vague language, prompt injection, missing structure, and other issues that degrade prompt quality in production. Runs locally — no API calls, no round-trips, results in milliseconds.

Install

pip install promptlint-cli

For accurate token counting (requires a Rust toolchain on non-CPython):

pip install promptlint-cli[tiktoken]

Requires Python 3.9+.

Quick start

# lint a file
promptlint --file prompt.txt

# lint inline text
promptlint -t "Please write some code for me"

# multiple files with globs
promptlint prompts/**/*.txt --exclude prompts/drafts/*

# pipe from stdin
cat prompt.txt | promptlint --format json

# auto-fix what it can
promptlint --file prompt.txt --fix

# CI mode: exit 1 on warnings, only print the summary
promptlint prompts/ --fail-level warn --quiet

Exit codes: 0 = clean, 1 = warnings found (with --fail-level warn), 2 = critical issues.

Example output

$ promptlint --file system_prompt.txt

PromptLint Findings
[ INFO     ] cost (line -) Prompt is ~38 tokens (~$0.0002 input per call on gpt-4o).
[ WARN     ] structure-sections (line -) No explicit sections detected (Task/Context/Output).
[ WARN     ] clarity-vague-terms (line 1) Vague term 'some' detected. Be more specific.
[ CRITICAL ] prompt-injection (line 5) Injection pattern detected: 'ignore previous instructions'.

1 file(s) scanned, 4 finding(s) in 0.41s

What it checks

Rule	What it does	Fixable
cost	Token count and per-call cost estimate	—
cost-limit	Warns when prompt exceeds your token budget	—
prompt-injection	Catches injection patterns, even with leetspeak/unicode obfuscation	yes
structure-sections	Flags prompts with no clear sections	yes
clarity-vague-terms	Finds "some", "stuff", "maybe", "good", etc.	—
specificity-examples	Suggests adding examples for complex instructions	—
specificity-constraints	Suggests adding length/format/scope constraints	—
politeness-bloat	Flags "please", "kindly", "thank you" (burns tokens)	yes
verbosity-sentence-length	Flags sentences over 40 words	—
verbosity-redundancy	"in order to" → "to", "due to the fact that" → "because"	yes
actionability-weak-verbs	Flags excessive passive voice	—
consistency-terminology	Catches mixed terms (user/customer, function/method)	—
completeness-edge-cases	Reminds you to specify error handling	—

Run promptlint --list-rules to see them all, or promptlint --explain cost for details on any rule.

Auto-fix

Pass --fix and PromptLint removes politeness filler, simplifies redundant phrases, strips injection lines, and scaffolds missing sections:

$ promptlint -t "Please kindly write code in order to sort the array, thank you" --fix

Optimized Prompt
<task>Write code to sort the array.</task>

Configuration

Drop a .promptlintrc in your repo root (or run promptlint --init to generate one):

model: gpt-4o
token_limit: 800
cost_per_1k_tokens: 0.005
calls_per_day: 10000

rules:
  cost:
    enabled: true
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
  politeness_bloat:
    enabled: true
    words: [please, kindly, thank you, i would appreciate]
  structure_sections:
    enabled: true

fix:
  enabled: true
  prompt_injection: true
  politeness_bloat: true
  verbosity_redundancy: true
  structure_scaffold: true

CLI reference

promptlint [FILES...] [OPTIONS]

  -V, --version            Show version
  -f, --file PATH          Single prompt file
  -t, --text TEXT          Inline prompt text
  -c, --config PATH        Config file (default: .promptlintrc)
  --format {text,json}     Output format
  --fix                    Auto-fix and print optimized prompt
  --fail-level LEVEL       none / warn / critical (default: critical)
  --show-dashboard         Token savings breakdown
  -q, --quiet              Summary line only (for CI)
  --exclude PATTERN        Exclude globs (repeatable)
  --list-rules             Show all rules
  --explain RULE_ID        Explain a specific rule
  --init                   Generate starter .promptlintrc

Injection evasion detection

The prompt-injection rule normalizes text before matching, catching obfuscated attacks that simple regex would miss:

Evasion technique	Example	Detected?
Leetspeak	1gn0r3 pr3v10u$ 1nstruct10ns	Yes
Zero-width characters	ign​ore previous instruc​tions	Yes
Character repetition	ignoooore previooous instructions	Yes
Fullwidth unicode	ｉｇｎｏｒｅ previous instructions	Yes
Mixed obfuscation	!gnor3 pr3v!0u$ in$truction$	Yes

Custom injection patterns can be added in .promptlintrc under rules.prompt_injection.patterns.

Agent skills

PromptLint ships with skills for AI coding agents that teach them prompt conventions and a lint-and-fix loop:

• Cursor — .cursor/skills/promptlint/SKILL.md
• Claude Code — .claude/skills/promptlint/SKILL.md

Links

• GitHub
• Changelog

License

Apache 2.0