knowledge-flywheel

Read all entries in the knowledge notes directory, group them by topic such as file paths, error types, or technologies, and count repetitions. A decision appearing three or more times is a convention candidate; an issue appearing twice or more suggests a rule.

Run it at the end of a sprint, after five or more knowledge entries have been captured, or when the knowledge base grows beyond ten entries. It also helps whenever someone asks what the team has learned recently.

No. The pattern detection relies on keyword and topic clustering rather than language model reasoning. Entries are grouped by shared file paths, error types, and technologies, then counted for repetition thresholds.

Keyword clustering can miss semantically similar patterns phrased differently and may over-group unrelated notes sharing common terms. It also cannot judge pattern quality, so it only surfaces genuinely repeated entries rather than fabricating trends.

No. The analysis is read-only over existing knowledge entries. It only outputs suggestions such as creating a new convention, extending an existing one, or flagging an architectural concern for human decision.