ig-learn

igap sync               # pull insights for posts younger than 7 days
igap learn              # rebuild ~/.instagram-autopilot/playbook.json and print it
igap report             # print the last playbook + recent posts

The scheduler runs sync + learn on every tick, so this is mostly for reading and explaining.

How it decides

• Outlier multiple = a post's views ÷ the account's own median. Raw views mostly measure follower count; the multiple measures the post.
• Watch-through = average watch time ÷ video length (Reels only).
• Numeric features (length, first cut, longest still, lead silence, loudness, caption length, hashtags, hour) are rank-correlated (Spearman) with both. |rho| ≥ 0.4 is reported as a signal.
• Buckets: length band, weekday+hour slot, and the hook tag set on each item.
• Confidence: collecting below learn.min_posts (6) settled posts, then low < 12, medium < 30, high. Nothing is applied while collecting.

What it changes on its own: target_max_s (best length band, if it beats the median by 20% with 2+ posts), max_shot_s (when still stretches hurt), max_lead_silence_s (when dead air hurts), and the posting slots. Everything else is advice for the user.

Talking about results

Be honest about sample size: 8 posts is a hint, not proof. Recommend changing one thing at a time (tag hooks with --hook so they can be compared) and say which signal a recommendation rests on. A post with a high multiple and low watch-through had a good hook and a weak middle; the reverse had a weak hook. Point at the specific timing feature when you can.