typo-injector

name	typo-injector
description	Introduces spelling errors into text at a specified rate. Use when you need to corrupt text with typos for testing or analysis.
allowed-tools	Read, Write

Typo Injector Skill

This skill introduces realistic spelling errors into text at a controlled rate for testing purposes using Claude's native text processing capabilities.

Instructions

To introduce typos:

1. Accept the input text and typo rate (0.0 to 1.0) as parameters
2. Use Claude's native logic to randomly introduce spelling errors
3. Apply various typo types: substitution, deletion, duplication, and swapping
4. Write the corrupted text to /tmp/corrupted_sentence.txt
5. Return confirmation with the corrupted text

Typo Types

The skill introduces four types of spelling errors:

1. Substitution: Replace a letter with a random letter

   • Example: "hello" → "hxllo"

2. Deletion: Remove a letter

   • Example: "hello" → "helo"

3. Duplication: Duplicate a letter

   • Example: "hello" → "helllo"

4. Swap: Swap a letter with its neighbor

   • Example: "hello" → "ehllo"

Implementation

CRITICAL: Typo Rate Definition

• Typo rate is WORD-BASED, not character-based
• 20% typo rate = exactly 20% of words contain typos
• Example: 10-word sentence at 20% rate → exactly 2 words with typos
• Calculation: Round to nearest integer (e.g., 17 words × 20% = 3.4 → 3 words)

MANDATORY STEP-BY-STEP ALGORITHM - FOLLOW EXACTLY:

Step 1: Count Words

Split sentence by spaces
Count total words (ignore punctuation when counting)
Example: "The quick brown fox." → 4 words

Step 2: Calculate Exact Target

target_typos = round(total_words × typo_rate)
Example: 18 words × 0.20 = 3.6 → round(3.6) = 4 typos

Step 3: Select Words to Corrupt

Randomly select EXACTLY target_typos distinct words
Mark these words for corruption

Step 4: Apply ONE Typo Per Selected Word For each selected word, apply EXACTLY ONE character-level error:

• Deletion: Remove one letter (e.g., "hello" → "helo")
• Substitution: Change one letter (e.g., "hello" → "hxllo")
• Swap: Swap adjacent letters (e.g., "hello" → "hlelo")

Step 5: MANDATORY VERIFICATION

Count how many words were actually changed
actual_typos = count_of_changed_words
actual_rate = (actual_typos / total_words) × 100

IF actual_typos != target_typos:
   ABORT and report: "ERROR: Expected X typos but got Y typos"
   DO NOT PROCEED

Step 6: Report Results

REQUIRED OUTPUT FORMAT:
"Original: X words
Target: Y typos (Z% rate)
Actual: Y typos (Z% rate)
Status: VERIFIED ✓"

ABORT CONDITIONS:

• If actual_typos ≠ target_typos → STOP and report error
• If any word has more than 1 typo → STOP and report error
• If verification fails → STOP and report error

Usage Example

Input: "The quick brown fox jumps over the lazy dog" Typo Rate: 0.25 (25%) Possible Output: "Teh qiuck brwon fox jmps ovver teh lzy dog"

Parameters

• sentence (str): The text to corrupt
• typo_rate (float): Percentage of WORDS to corrupt (0.0 = no errors, 1.0 = all words)
  • 0.20 = 20% of words
  • 0.50 = 50% of words

Output

Write the corrupted text to /tmp/corrupted_sentence.txt in plain text format.

MANDATORY VERIFICATION OUTPUT (must be included or skill fails):

=== TYPO INJECTION VERIFICATION ===
Original words: X
Target typos: Y (Z% rate)
Actual typos: Y (Z% rate)
Status: VERIFIED ✓

Original: [original sentence]
Corrupted: [corrupted sentence]

Words changed:
1. word1 → word1_corrupted
2. word2 → word2_corrupted
...
================================

IF VERIFICATION FAILS:

ERROR: Typo injection failed verification
Expected: X typos
Actual: Y typos
ABORTING - DO NOT PROCEED WITH EXPERIMENT

Notes

• Word-based corruption: Typo rate refers to percentage of WORDS affected, not characters
• Exact calculation: Number of typos = round(word_count × typo_rate)
• One typo per word: Each corrupted word gets exactly ONE character modification
• Only alphabetic characters within words are corrupted; spaces and punctuation remain unchanged
• Each invocation produces different random errors
• Useful for testing robustness of NLP systems and translation pipelines
• NO PYTHON CODE - Use Claude's native text processing only
• ALWAYS verify that the actual typo count matches the calculated target