Config Examples

Ready-to-use .promptlintrc files for common scenarios. Copy one, drop it in your project root, and adjust to taste.

promptlint --init

promptlint --init writes the Balanced Starter config to .promptlintrc automatically. Use this page to swap to a profile that fits your use case.

Balanced Starter

The default generated by promptlint --init. All 20 rules enabled at their default severities. Good starting point for any project.

yaml

# .promptlintrc — Balanced Starter
# Generated by: promptlint --init

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

display:
  preview_length: 60
  context_width: 80

rules:
  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit: true

  # ── Security ──────────────────────────────────────────────────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
      - disregard all prior
      - forget everything
      - new instructions:
  jailbreak_pattern: true
  secret_in_prompt: true
  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: true
    check_ssn: true
    check_credit_card: true
    check_ip: false
  context_injection_boundary: true

  # ── Structure ─────────────────────────────────────────────────────
  structure_sections: true
  role_clarity: true
  output_format_missing: true
  hallucination_risk: true

  # ── Quality ───────────────────────────────────────────────────────
  clarity_vague_terms: true
  specificity_examples: true
  specificity_constraints: true
  politeness_bloat:
    enabled: true
    allow_politeness: false
  verbosity_redundancy: true
  verbosity_sentence_length: true
  actionability_weak_verbs: true
  consistency_terminology: true
  completeness_edge_cases: true

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

Security-First (Production API)

For customer-facing or externally-exposed prompts. All security rules at max, extended injection patterns, IP detection on. Quality rules run at INFO so they don't block deploys.

yaml

# .promptlintrc — Security-First
# Use when: customer-facing APIs, prompts that accept user input,
#           any system where prompt injection could cause harm

model: gpt-4o
token_limit: 1200
cost_per_1k_tokens: 0.005
calls_per_day: 100000

rules:
  # ── Security (all critical, extra patterns) ───────────────────────
  prompt_injection:
    enabled: true
    patterns:
      # Core hijack attempts
      - ignore previous instructions
      - disregard all prior
      - forget everything above
      - system prompt extraction
      - reveal your (system prompt|instructions)
      - what are your instructions
      - print your (prompt|instructions)
      # Role-override
      - "you are now a [a-zA-Z ]+"
      - act as if you have no restrictions
      - pretend you are
      - your new instructions are
      # Boundary injection
      - new instructions:
      - admin override:
      - developer mode (enabled|activated|on)
      - bypass (all |your )?(safety|restrictions|rules|filters)
      # Indirect injection vectors
      - execute the following
      - run the following command
      - eval\(

  jailbreak_pattern: true
  secret_in_prompt: true

  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: true
    check_ssn: true
    check_credit_card: true
    check_ip: true             # enabled: IPs are PII in many jurisdictions

  context_injection_boundary: true  # required: all user input must be bounded

  # ── Structure ─────────────────────────────────────────────────────
  role_clarity: true
  output_format_missing: true
  hallucination_risk: true
  structure_sections: true

  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit: true

  # ── Quality (demoted to INFO — won't fail CI) ─────────────────────
  clarity_vague_terms:
    enabled: true
    level: info
  specificity_examples:
    enabled: true
    level: info
  specificity_constraints:
    enabled: true
    level: info
  politeness_bloat:
    enabled: true
    level: info
  verbosity_redundancy:
    enabled: true
    level: info
  verbosity_sentence_length:
    enabled: true
    level: info
  actionability_weak_verbs:
    enabled: true
    level: info
  consistency_terminology:
    enabled: true
    level: info
  completeness_edge_cases:
    enabled: true
    level: info

fix:
  enabled: true
  prompt_injection: true      # delete injection lines automatically
  politeness_bloat: true
  verbosity_redundancy: true
  structure_scaffold: false   # don't auto-scaffold: you control structure manually

CI Gate (Strict)

Block merges that introduce security issues or major quality regressions. All WARN+ rules are hard failures. Designed for a CI pipeline that runs promptlint --fail-level warn.

yaml

# .promptlintrc — CI Gate
# Use when: enforcing quality in a mono-repo, blocking PR merges,
#           prompt review in a team environment
# Run as:  promptlint --file "prompts/**/*.txt" --fail-level warn --format sarif

model: gpt-4o
token_limit: 600          # tight budget: prompts must stay lean
cost_per_1k_tokens: 0.005
calls_per_day: 50000

rules:
  # ── Security (CRITICAL — always blocks) ──────────────────────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
      - disregard all prior
      - forget everything
      - new instructions:
      - admin override:
      - bypass restrictions
  jailbreak_pattern: true
  secret_in_prompt: true
  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: true
    check_ssn: true
    check_credit_card: true
    check_ip: false
  context_injection_boundary: true

  # ── Structure (WARN — blocks on --fail-level warn) ────────────────
  role_clarity: true
  output_format_missing: true
  hallucination_risk: true
  structure_sections: true

  # ── Quality (promoted to WARN to enforce team standards) ──────────
  clarity_vague_terms: true           # default WARN — blocks
  specificity_examples:
    enabled: true
    level: warn                       # promoted from INFO
  specificity_constraints:
    enabled: true
    level: warn                       # promoted from INFO
  completeness_edge_cases:
    enabled: true
    level: warn                       # promoted from INFO
  actionability_weak_verbs:
    enabled: true
    level: warn                       # promoted from INFO

  # ── Quality (stay at INFO — advisory only) ────────────────────────
  politeness_bloat: true
  verbosity_redundancy: true
  verbosity_sentence_length: true
  consistency_terminology: true

  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit: true                    # CRITICAL above token_limit

fix:
  enabled: false    # CI runs lint only; developers fix locally with --fix

Cost Optimization

Focus on token reduction. Strips politeness, collapses redundancy, and enforces a hard token budget. Useful for high-volume inference where every token saved matters.

yaml

# .promptlintrc — Cost Optimization
# Use when: high call volume (>10k/day), tight token budget,
#           optimizing GPT-4o/Claude costs at scale

model: gpt-4o
token_limit: 400            # aggressive: keep system prompts very lean
cost_per_1k_tokens: 0.005
calls_per_day: 100000       # project real-scale cost impact

rules:
  # ── Cost (primary focus) ──────────────────────────────────────────
  cost: true
  cost_limit: true

  # ── Token bloat rules (promoted to WARN to force fixing) ──────────
  politeness_bloat:
    enabled: true
    allow_politeness: false   # WARN severity — blocks CI
    words:
      - please
      - kindly
      - i would appreciate
      - thank you
      - be so kind as to
      - if possible
      - as needed
      - as appropriate
    savings_per_hit: 1.5

  verbosity_redundancy:
    enabled: true             # INFO — advisory
  verbosity_sentence_length:
    enabled: true             # INFO — advisory

  # ── Security (keep: no cost to these) ────────────────────────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
  jailbreak_pattern: true
  secret_in_prompt: true
  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: false   # reduce false positives in cost-focused context
    check_ssn: true
    check_credit_card: true
    check_ip: false
  context_injection_boundary: true

  # ── Quality (disable or demote: focus is cost, not style) ─────────
  structure_sections:
    enabled: true
    level: info
  role_clarity:
    enabled: true
    level: info
  output_format_missing: true     # keep at WARN: undefined format → more tokens
  clarity_vague_terms:
    enabled: true
    level: info
  specificity_examples: false     # not cost-relevant
  specificity_constraints: false  # not cost-relevant
  completeness_edge_cases: false  # not cost-relevant
  actionability_weak_verbs: false
  consistency_terminology: false
  hallucination_risk:
    enabled: true
    level: info

fix:
  enabled: true
  politeness_bloat: true    # auto-strip immediately
  verbosity_redundancy: true  # auto-collapse immediately
  prompt_injection: true
  structure_scaffold: false   # don't add wrapper (adds tokens)
  normalize_spacing: true

Customer Support Bot

Strict on injection and PII (customer data in context), soft on style (support tone is intentionally warm). Hallucination rule tuned since answers come from a knowledge base.

yaml

# .promptlintrc — Customer Support Bot
# Use when: support chatbots, help desk agents, FAQ bots,
#           any prompt that receives customer-provided input

model: gpt-4o
token_limit: 1500           # support prompts carry product context
cost_per_1k_tokens: 0.005
calls_per_day: 20000

rules:
  # ── Security (strict: customer data is in-scope) ──────────────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - forget everything above
      - system prompt extraction
      - reveal your (system prompt|instructions)
      - "you are now a [a-zA-Z ]+"
      - bypass (all |your )?(safety|restrictions|rules|filters)
      - disregard all prior
      - new instructions:
      - act as if you have no restrictions
  jailbreak_pattern: true
  secret_in_prompt: true

  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: true
    check_ssn: true
    check_credit_card: true
    check_ip: false

  context_injection_boundary: true    # required: {{USER_MESSAGE}} must be bounded

  # ── Structure ─────────────────────────────────────────────────────
  role_clarity: true
  output_format_missing: true
  hallucination_risk: false   # disabled: answers grounded in KB context
  structure_sections: true

  # ── Quality (relaxed: support tone is intentionally warm) ─────────
  politeness_bloat:
    enabled: true
    allow_politeness: true    # INFO severity: warm language is intentional
  actionability_weak_verbs:
    enabled: true
    level: info               # "feel free to ask" is acceptable in support
  clarity_vague_terms:
    enabled: true
    level: info
  specificity_examples:
    enabled: true
    level: info
  specificity_constraints:
    enabled: true
    level: info
  verbosity_redundancy:
    enabled: true
    level: info
  verbosity_sentence_length:
    enabled: true
    level: info
  completeness_edge_cases: true
  consistency_terminology: true

  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit: true

fix:
  enabled: true
  prompt_injection: true
  politeness_bloat: false     # don't strip: warm language is intentional
  verbosity_redundancy: true
  structure_scaffold: true

Code Generation

Strict on structure, completeness, and constraints — a code generation prompt without edge-case handling and output format will produce inconsistent results. Extends injection patterns for code-specific attack vectors.

yaml

# .promptlintrc — Code Generation
# Use when: code synthesis, refactoring assistants, test generation,
#           any prompt that produces executable output

model: gpt-4o
token_limit: 1000
cost_per_1k_tokens: 0.005
calls_per_day: 10000

rules:
  # ── Security (extended for code contexts) ─────────────────────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
      - disregard all prior
      - forget everything
      - new instructions:
      - admin override:
      # Code-specific injection vectors
      - "eval\\("                  # eval injection
      - "exec\\("                  # exec injection
      - "__import__"               # Python import injection
      - "subprocess\\."            # subprocess in output
      - "os\\.system"              # shell injection
      - "rm -rf"
      - "DROP TABLE"
      - ";.*--"                    # SQL comment injection
  jailbreak_pattern: true
  secret_in_prompt: true
  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: false
    check_ssn: true
    check_credit_card: true
    check_ip: false
  context_injection_boundary: true

  # ── Structure (strict: code prompts need clear specs) ─────────────
  role_clarity: true
  output_format_missing: true    # always specify: Python? TypeScript? plain text?
  structure_sections: true
  hallucination_risk: true       # "write a function using library X" is factual

  # ── Quality (strict: code prompts must be precise) ────────────────
  clarity_vague_terms: true
  specificity_examples:
    enabled: true
    level: warn                  # promoted: examples are critical for code gen
  specificity_constraints:
    enabled: true
    level: warn                  # promoted: constraints define correct output
  completeness_edge_cases:
    enabled: true
    level: warn                  # promoted: edge cases are critical for code
  actionability_weak_verbs:
    enabled: true
    level: warn                  # no "try to" in code: be exact
  consistency_terminology:
    enabled: true
    custom_term_pairs:
      - ["function", "method", "procedure"]
      - ["error", "exception", "fault"]
      - ["return", "output", "result", "response"]
      - ["parameter", "argument", "input"]
  verbosity_redundancy: true
  verbosity_sentence_length: true

  # ── Politeness (off: code prompts are instructions, not requests) ──
  politeness_bloat:
    enabled: true
    allow_politeness: false

  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit: true

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

RAG / Retrieval-Augmented Generation

Prompts that inject retrieved context at runtime. Hallucination risk is suppressed (grounding is structural), injection boundary is required for every template variable.

yaml

# .promptlintrc — RAG / Grounded Prompts
# Use when: RAG pipelines, document Q&A, search-augmented generation,
#           any prompt with {{CONTEXT}} or {{RETRIEVED_DOCS}} slots

model: gpt-4o
token_limit: 2000           # RAG prompts carry retrieved documents
cost_per_1k_tokens: 0.005
calls_per_day: 25000

rules:
  # ── Security ──────────────────────────────────────────────────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
      - disregard all prior
      - forget everything
      - new instructions:
      - bypass (all |your )?(safety|restrictions|rules|filters)
      # RAG-specific: injections hidden inside retrieved documents
      - "\\[INST\\]"               # Llama-style instruction tag in docs
      - "<\\|system\\|>"           # system tag appearing in retrieved text
      - "END_OF_SYSTEM_PROMPT"
  jailbreak_pattern: true
  secret_in_prompt: true
  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: true
    check_ssn: true
    check_credit_card: true
    check_ip: false

  context_injection_boundary: true  # required: every {{VAR}} must be bounded

  # ── Structure ─────────────────────────────────────────────────────
  role_clarity: true
  output_format_missing: true
  structure_sections: true
  hallucination_risk: false         # disabled: retrieval is the grounding

  # ── Quality ───────────────────────────────────────────────────────
  clarity_vague_terms: true
  specificity_examples: false       # disabled: RAG prompts use templates, not examples
  specificity_constraints: true
  completeness_edge_cases:
    enabled: true                   # handle: empty retrieval, conflicting docs, low score
  consistency_terminology: true
  politeness_bloat: true
  verbosity_redundancy: true
  verbosity_sentence_length: true
  actionability_weak_verbs: true

  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit: true

fix:
  enabled: true
  prompt_injection: true
  politeness_bloat: true
  verbosity_redundancy: true
  structure_scaffold: false   # RAG prompts have custom structure; don't overwrite

Creative Writing / Soft Instructions

For open-ended generative tasks where hedged language and vague terms are intentional. Security rules remain active — injection attacks happen in creative contexts too.

yaml

# .promptlintrc — Creative Writing
# Use when: story generation, brainstorming, open-ended content creation,
#           prompts where "feel free to" and "consider" are intentional

model: gpt-4o
token_limit: 2000
cost_per_1k_tokens: 0.005
calls_per_day: 1000

rules:
  # ── Security (always on) ──────────────────────────────────────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
      - disregard all prior
      - new instructions:
  jailbreak_pattern: true
  secret_in_prompt: true
  pii_in_prompt:
    enabled: true
    check_email: true
    check_phone: true
    check_ssn: true
    check_credit_card: true
    check_ip: false
  context_injection_boundary: true

  # ── Structure (relaxed: creative prompts are less structured) ─────
  role_clarity:
    enabled: true
    level: info             # advisory: personas help but aren't mandatory
  output_format_missing:
    enabled: true
    level: info             # advisory: format often left open intentionally
  structure_sections:
    enabled: true
    level: info
  hallucination_risk:
    enabled: true
    level: info             # creative prompts often ask for imagined content

  # ── Quality (disabled: soft language is intentional) ──────────────
  actionability_weak_verbs: false   # "consider", "feel free" are intentional
  clarity_vague_terms: false        # "something interesting" is valid creative direction
  specificity_examples: false       # open-ended prompts don't need examples
  specificity_constraints: false    # constraints kill creative output
  completeness_edge_cases: false    # edge cases don't apply to creative tasks
  consistency_terminology: false    # intentional variety enriches creative prompts

  politeness_bloat:
    enabled: true
    allow_politeness: true          # INFO only: polite framing is acceptable
  verbosity_redundancy: true        # still useful: removes pure waste
  verbosity_sentence_length:
    enabled: true
    level: info
  actionability_weak_verbs: false

  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit: true

fix:
  enabled: true
  prompt_injection: true
  politeness_bloat: false           # keep warm language
  verbosity_redundancy: true
  structure_scaffold: false         # don't impose structure on creative prompts

Internal / Development (Relaxed)

For prompts used in internal tools, development environments, or personal projects where iteration speed matters more than strict enforcement.

yaml

# .promptlintrc — Internal / Development
# Use when: internal tooling, developer environments, rapid prototyping,
#           personal projects where strict gates slow down iteration

model: gpt-4o
token_limit: 2000           # generous limit for dev exploration
cost_per_1k_tokens: 0.005
calls_per_day: 500          # low volume: cost projection less relevant

rules:
  # ── Security (keep critical, relax others to INFO) ────────────────
  prompt_injection:
    enabled: true           # always: injection affects dev systems too
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
  jailbreak_pattern: true
  secret_in_prompt: true    # always: dev prompts accidentally embed real keys
  pii_in_prompt:
    enabled: true
    check_email: true       # dev prompts often use real test emails
    check_phone: false
    check_ssn: true
    check_credit_card: true
    check_ip: false
  context_injection_boundary:
    enabled: true
    level: info             # advisory in dev: boundary not always needed

  # ── Structure (all advisory) ──────────────────────────────────────
  role_clarity:
    enabled: true
    level: info
  output_format_missing:
    enabled: true
    level: info
  structure_sections:
    enabled: true
    level: info
  hallucination_risk:
    enabled: true
    level: info

  # ── Quality (all advisory — iterating quickly) ────────────────────
  clarity_vague_terms:
    enabled: true
    level: info
  specificity_examples:
    enabled: true
    level: info
  specificity_constraints:
    enabled: true
    level: info
  politeness_bloat:
    enabled: true
    allow_politeness: true  # INFO: don't interrupt typing flow
  verbosity_redundancy: true
  verbosity_sentence_length:
    enabled: true
    level: info
  actionability_weak_verbs:
    enabled: true
    level: info
  consistency_terminology:
    enabled: true
    level: info
  completeness_edge_cases:
    enabled: true
    level: info

  # ── Cost ──────────────────────────────────────────────────────────
  cost: true
  cost_limit:
    enabled: true
    level: warn             # advisory: dev prompts can be longer

fix:
  enabled: true
  prompt_injection: true
  politeness_bloat: false   # don't auto-strip in dev
  verbosity_redundancy: true
  structure_scaffold: true

Non-English Prompts

Security rules work in any language (regex-based). Quality rules are English-optimized and will generate false positives on other languages — disable them entirely.

yaml

# .promptlintrc — Non-English Prompts
# Use when: prompts written in French, Spanish, German, Japanese, etc.
# Why: quality rules (vague terms, politeness, redundancy, weak verbs)
#      use English patterns and have high false-positive rates on other languages.
#      Security rules are regex-based and language-agnostic.

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

rules:
  # ── Security (all on: regex patterns, language-agnostic) ──────────
  prompt_injection:
    enabled: true
    patterns:
      - ignore previous instructions
      - system prompt extraction
      - "you are now a [a-zA-Z ]+"
      - disregard all prior
      - forget everything
      - new instructions:
      # Add language-specific patterns if needed, e.g.:
      # - ignorez les instructions précédentes   # French
      # - ignorar instrucciones anteriores        # Spanish
      # - Vergiss alle vorherigen Anweisungen     # German
  jailbreak_pattern: true
  secret_in_prompt: true
  pii_in_prompt:
    enabled: true
    check_email: true       # email format is language-agnostic
    check_phone: true       # phone regex is international
    check_ssn: true
    check_credit_card: true
    check_ip: false
  context_injection_boundary: true

  # ── Cost (language-agnostic) ──────────────────────────────────────
  cost: true
  cost_limit: true

  # ── Structure (partially useful) ──────────────────────────────────
  structure_sections:
    enabled: true           # XML/JSON structure is language-agnostic
    level: info
  role_clarity: false       # disabled: role phrase list is English-only
  output_format_missing: false  # disabled: format keywords list is English-only
  hallucination_risk: false     # disabled: factual question patterns are English

  # ── Quality (all disabled: English patterns, high false positives) ─
  clarity_vague_terms: false
  specificity_examples: false
  specificity_constraints: false
  politeness_bloat: false
  verbosity_redundancy: false
  verbosity_sentence_length: false
  actionability_weak_verbs: false
  consistency_terminology: false
  completeness_edge_cases: false

fix:
  enabled: true
  prompt_injection: true
  politeness_bloat: false     # disabled: words list is English
  verbosity_redundancy: false # disabled: phrase list is English
  structure_scaffold: false

Quick Reference

Use case	Key changes from default
Balanced Starter	All 20 rules, default severities
Security-First	Extended patterns, IP on, quality → INFO
CI Gate	Quality rules → WARN, fix disabled
Cost Optimization	Token limit 400, bloat rules → WARN, fix on
Customer Support	Warm language allowed, hallucination off
Code Generation	Edge cases + constraints → WARN, code injection patterns
RAG	Hallucination off, injection boundary required, examples off
Creative Writing	Style rules off, security on
Internal Dev	All quality → INFO, generous token limit
Non-English	All quality rules off, security rules only

Config Examples ​

Balanced Starter ​

Security-First (Production API) ​

CI Gate (Strict) ​

Cost Optimization ​

Customer Support Bot ​

Code Generation ​

RAG / Retrieval-Augmented Generation ​

Creative Writing / Soft Instructions ​

Internal / Development (Relaxed) ​

Non-English Prompts ​

Quick Reference ​

Config Examples

Balanced Starter

Security-First (Production API)

CI Gate (Strict)

Cost Optimization

Customer Support Bot

Code Generation

RAG / Retrieval-Augmented Generation

Creative Writing / Soft Instructions

Internal / Development (Relaxed)

Non-English Prompts

Quick Reference