Extract Error Patterns

Core Position#

This skill analyzes unstructured or semi-structured data and extracts structured information according to defined patterns. It is NOT a simple search — it applies intelligent pattern matching and classification.

Key responsibilities:

• Parse input data (logs/text/code/etc.) and identify structural elements
• Apply regex patterns and heuristic rules to extract relevant information
• Classify and categorize extracted items with confidence scores
• Handle ambiguous cases by reporting uncertainty rather than guessing

Modes#

/extract-error-patterns --verbose#

Verbose mode. Includes confidence scores, pattern matches, and edge case details.

/extract-error-patterns --summary#

Summary mode. Returns only high-confidence extractions with a count summary.

Execution Steps#

1. Parse input — Read the input data; detect format (log file, JSON, plain text, etc.)

   • If input is empty or unreadable, report: "Input is empty or unreadable"

2. Identify structure — Find delimiters, sections, and repeating patterns in the data

   • If no structure detected, treat as plain text and proceed with text-based extraction

3. Apply extraction patterns — Run regex/heuristic patterns; collect all matches with positions

   • Track: matched_pattern, matched_text, position, context (surrounding lines)

4. Classify results — Categorize extractions by type/severity; assign confidence score (0-1)

   • Low confidence (< 0.6) items should be flagged for manual review

5. Deduplicate — Remove exact duplicates; flag near-duplicates (similar within 90% match)

6. Report — Return structured list with: type, value, confidence, position, context

Mandatory Rules#

Do not#

• Do not suppress extraction failures — if a section cannot be parsed, report it separately
• Do not assign high confidence to extractions from ambiguous or inconsistent patterns
• Do not extract personal data (PII) without explicit user confirmation and data handling rules
• Do not apply extraction patterns that were not explicitly defined or reviewed
• Do not assume encoding — always validate and report the detected input encoding

Do#

• Report total extraction count broken down by type and confidence level
• Include the original position/context for every extraction so the user can verify
• Flag any extraction that spans multiple records or has unusual characteristics
• Handle inputs up to the documented size limit; if exceeded, report and truncate
• Log all patterns applied and the match count for each pattern

Quality Bar#

CriterionMinimumIdealExtraction coverage>= 95% of identifiable entities100% with confidence score per itemFalse positive rate< 5% of extractions are wrong< 1% with manual review flaggingConfidence calibrationScore 0-1, low < 0.6 flaggedAll low-confidence items reviewed and re-scoredContext preservedEvery extraction has source locationSource location + surrounding context linesPattern documentationEvery pattern has a documented purposePatterns rated by precision/recall tradeoffA good extraction result contains confidence scores and source locations for every item.

Good vs. Bad Examples#

ScenarioBadGoodLow confidenceMarks as high confidence, flags no reviewReports "Extraction [X]: confidence 0.4 — flagged for manual review"Ambiguous inputForces extraction, guessesReports "Cannot classify [X] — insufficient context, skipped"PII detectedExtracts without warningStops and asks: "PII detected in [location] — confirm before extracting"Pattern missSilent failure, returns emptyReports "No patterns matched input — check format or add custom pattern"Large inputLoads all, crashesStreams, reports "Processed 10K lines, found 142 matches (truncated at limit)"