Skip to main content
Auto-Skill’s pattern detection is highly configurable. You can customize detection thresholds, sequence lengths, lookback periods, and ignored tools to match your workflow.

​
Detection Configuration

Pattern detection behavior is controlled by the DetectionConfig interface:
DetectionConfig Interface
export interface DetectionConfig {
  minOccurrences: number;        // Minimum times a pattern must repeat
  minSequenceLength: number;     // Minimum number of tools in a sequence
  maxSequenceLength: number;     // Maximum number of tools in a sequence
  lookbackDays: number;          // How far back to analyze sessions
  minConfidence: number;         // Minimum confidence score (0.0-1.0)
  ignoredTools: string[];        // Tools to exclude from pattern detection
}

​
Default Values

minOccurrences

Default: 3A tool sequence must appear at least 3 times across different sessions to be considered a pattern.

minSequenceLength

Default: 2Sequences must contain at least 2 tools. Single-tool patterns are too generic.

maxSequenceLength

Default: 10Sequences longer than 10 tools are truncated. Very long patterns are often noise.

lookbackDays

Default: 7Analyzes the last 7 days of sessions. Older patterns may be stale.

minConfidence

Default: 0.7Only patterns with 70%+ confidence are suggested for skill generation.

ignoredTools

Default: ["AskUserQuestion"]These tools are excluded from pattern detection (too interactive).

​
Configuration File

You can override defaults using a project-local configuration file:
.claude/auto-skill.local.md
---
detection:
  min_occurrences: 5
  min_sequence_length: 3
  max_sequence_length: 8
  lookback_days: 14
  min_confidence: 0.8
  ignored_tools:
    - AskUserQuestion
    - WebFetch
    - Task
---

# Auto-Skill Local Configuration

This file customizes pattern detection for this project.
The configuration file uses YAML frontmatter, similar to SKILL.md files. The body content is optional and ignored.

​
Loading Configuration

Configuration is loaded with project-specific overrides:
Loading Config
import { loadConfig } from '@matrixy/auto-skill';

// Load default config
const defaultConfig = loadConfig();

// Load config with project overrides
const projectConfig = loadConfig('/path/to/project');
The loader:
  1. Starts with default values
  2. Looks for .claude/auto-skill.local.md in the project directory
  3. Parses YAML frontmatter
  4. Merges overrides into the default config
  5. Returns the merged Config object

​
Confidence Scoring

Auto-Skill calculates confidence using a weighted formula:
1

Occurrence Score (40%)

Logarithmic scaling rewards frequent patterns:
const occurrenceScore = Math.min(
  1.0,
  Math.log(occurrenceCount + 1) / Math.log(10)
);
  • 3 occurrences → ~0.60
  • 5 occurrences → ~0.78
  • 10 occurrences → ~1.00
2

Length Score (20%)

Optimal sequence length is 3-5 tools:
if (sequenceLength >= 3 && sequenceLength <= 5) {
  lengthScore = 1.0;
} else if (sequenceLength === 2) {
  lengthScore = 0.7;
} else if (sequenceLength > 5) {
  lengthScore = Math.max(0.5, 1.0 - (sequenceLength - 5) * 0.1);
}
3

Success Score (25%)

Ratio of successful executions:
const successScore = successCount / totalCount;
  • 100% success → 1.0
  • 80% success → 0.8
  • 50% success → 0.5
4

Recency Score (15%)

Recent patterns score higher:
const daysSinceLast = (nowMs - lastSeenMs) / 86_400_000;
const recencyScore = Math.max(0.5, 1.0 - daysSinceLast * 0.05);
  • Used today → 1.0
  • 5 days ago → 0.75
  • 10 days ago → 0.5 (floor)
Final confidence: 
const confidence =
  occurrenceScore * 0.4 +
  lengthScore * 0.2 +
  successScore * 0.25 +
  recencyScore * 0.15;

​
Ignoring Tools

Some tools should be excluded from pattern detection because they:
  • Are too interactive (AskUserQuestion)
  • Fetch external content (WebFetch)
  • Delegate to other agents (Task)
  • Are Auto-Skill commands themselves
detection:
  ignored_tools:
    - AskUserQuestion
    - WebFetch
    - Task
    - TodoWrite

​
Tuning for Your Workflow

Lower the occurrence threshold and minimum confidence:
detection:
  min_occurrences: 2
  min_confidence: 0.6
Tradeoff: You may see more false positives and noise.
Raise the confidence threshold and occurrence requirement:
detection:
  min_occurrences: 5
  min_confidence: 0.85
Tradeoff: You may miss useful patterns that haven’t repeated enough yet.
Increase the lookback period:
detection:
  lookback_days: 30
Tradeoff: Older patterns may no longer be relevant.
Increase the max sequence length:
detection:
  max_sequence_length: 15
Tradeoff: Very long sequences are harder to match and may be too specific.

​
Pattern Detection Pipeline

​
Advanced: Per-Tool Weights

Coming in v6.0: Per-tool weighting will allow you to assign importance scores to specific tools, influencing confidence calculations.
detection:
  tool_weights:
    Bash: 1.2      # Shell commands are high-signal
    Edit: 1.1      # Code edits are important
    Read: 0.8      # Reading files is common
    Grep: 0.9      # Searching is exploratory

​
Example Patterns by Confidence

{
  "toolSequence": ["Read", "Edit", "Bash"],
  "occurrenceCount": 8,
  "successRate": 1.0,
  "confidence": 0.92,
  "suggestedName": "read-and-run",
  "description": "Read code, edit it, then run tests"
}

​
Next Steps

Configuration Reference

Complete reference for all config options

Integration Guide

Learn how patterns are detected in real-time

Build docs developers (and LLMs) love

Get started for freeTalk to us