Skip to content

Skill Synthesis Settings

All settings live in ~/.ptah/settings.json under the skillSynthesis.* prefix. Edit through Settings → Skill Synthesis.

KeyDefaultWhat it does
skillSynthesis.enabledtrueMaster kill-switch — disables both detection and promotion
skillSynthesis.successesToPromote3Successful repetitions before a single candidate is promoted directly
skillSynthesis.suggestionMinClusterSize2Similar candidates required before a cluster is distilled into a Recommended skill
skillSynthesis.judgeEnabledtrueWhether the quality judge gates promotion and recommendations
skillSynthesis.minJudgeScore6.0Minimum average judge score (of 10, across five criteria) for a skill to pass
skillSynthesis.dedupCosineThreshold0.85Embedding similarity above which a trajectory is treated as a duplicate of an active skill
skillSynthesis.maxActiveSkills200Residency cap; the weakest resident is demoted to dormant (not deleted) when exceeded
skillSynthesis.curatorEnabledtrueWhether the periodic Curator pass runs (clustering, recommendations, auto-enhancement)
skillSynthesis.curatorIntervalHours24How often the Curator pass runs
skillSynthesis.candidatesDir''Override location for promoted skills; empty string means ~/.ptah/skills/
  • Lowering successesToPromote to 2 makes the direct-promotion path eager — more skills faster, including some you’d reject. Pair with active review.
  • Lowering suggestionMinClusterSize would let single sessions become recommendations; the default 2 is deliberate so Recommended only shows genuinely repeated workflows.
  • Lowering minJudgeScore lets weaker skills through; raising it makes Recommended stricter. The judge fails open, so an unavailable model never blocks the pipeline regardless of this value.
  • Raising dedupCosineThreshold to 0.92 allows near-duplicates through — useful for fine-grained variations as separate skills; usually the default is what you want.
  • maxActiveSkills is a soft governance cap. Demotion is to dormant, so nothing is lost when you exceed it.