knowledge-flywheel

Analyze captured knowledge for patterns — detect repeated decisions, suggest conventions, surface "you keep running into X" insights.

SKILL.md

GitHub

What I do

Analyze the knowledge base (.hero/knowledge/) for patterns that should be promoted to conventions, decisions, or warnings. This turns raw captured notes into structured team intelligence.

When to use me

  • At the end of a sprint or after 5+ knowledge entries have been captured
  • When /prime detects a growing knowledge base (>10 entries)
  • When the user asks "what have we learned?" or "any patterns?"

How to use me

  1. Read all entries in .hero/knowledge/notes/
  2. Group by topic (file paths mentioned, error types, technologies)
  3. Identify patterns:
    • Repeated decisions: same choice made 3+ times → suggest a convention
    • Recurring issues: same debugging finding 2+ times → suggest a rule
    • Converging practices: similar approaches across specs → suggest codifying
  4. For each pattern found, suggest an action:
    • /convention to codify a repeated practice
    • Add to an existing convention if one covers the topic
    • Flag as a potential architectural concern if it's a structural pattern

Output format

Knowledge flywheel — N entries analyzed

Patterns detected:

  1. Convention candidate: "Error handling in API routes"
     Seen in: 3 notes (auth-debugging, payment-retry, export-errors)
     Suggestion: /convention codify API error response retry pattern

  2. Recurring issue: "Database connection pool exhaustion"
     Seen in: 2 notes (monday-outage, load-test-findings)
     Suggestion: Add connection pool monitoring to operational runbook

  3. No new patterns (N entries too unique to cluster)

What not to do

  • Don't fabricate patterns — only surface what's genuinely repeated
  • Don't suggest conventions for one-off decisions
  • Don't modify existing knowledge entries
  • Don't require an LLM — pattern detection is keyword/topic clustering