AI Skill Suggestions

The scanner identifies repeating patterns in your conversation history and generates production-grade SKILL.md files using a cost-efficient two-stage pipeline.

Pipeline

Evidence Gating

Before calling any API, the scanner evaluates whether conversation history has enough signal for quality suggestions:

Check	Threshold	If Fails
Prompt count	≥10 prompts	Skip API calls entirely
Pattern strength	≥3 occurrences per cluster	Report as "emerging patterns" only
Cross-project	≥2 distinct projects	Report as "emerging patterns" only
Tool chain patterns	≥3 matching chains across ≥2 projects	Contributes to pass threshold

When evidence is weak, the report includes an "Emerging Patterns" section instead of full skill suggestions — no API tokens wasted.

Pre-Clustering Algorithm

1. Normalize prompts (lowercase, strip paths/URLs/hashes)
2. Extract 3-word leading phrase as cluster key
3. Group by key, record count + first 3 examples
4. Sort by count, keep top 20
5. Aggregate tool chain fingerprints (first 3 tools in each session)

Only clusters and chain patterns are sent to the API — not raw prompt history.

Model Fallback Chain

Stage	Primary	Fallback 1	Fallback 2
Identify	gemini-3.1-flash-lite-preview	gemini-3-flash-preview	—
Write	gemini-3.1-pro-preview	gemini-3-pro-preview	gemini-3-flash-preview

If all models fail for a candidate, the scanner returns a stub template with key tools listed.

Prerequisites

# Option A: Vertex AI (uses gcloud auth ADC)
export GOOGLE_CLOUD_PROJECT="your-project"

# Option B: API key
export GOOGLE_API_KEY="your-key"

Without either variable, the scanner runs but skips skill suggestions.

Quality Standards

Generated skills follow Agent Skills specification:

• Procedures over declarations
• Defaults over menus
• Exact tool calls with concrete parameters
• Gotchas sections from environment-specific traps
• Validation loops after every multi-step workflow
• Progressive disclosure (SKILL.md focused, deep detail in references/)