Orbitant Learning Insights

Overview

Centralized intelligence layer that mines past session data to improve current learning sessions. Not user-facing — invoked by /learn and /end-learn only.

Mode 1: Build Context

Called by /learn at session start, before content delivery.

What to Read

1. ~/.claude/learning/insights.md — persistent summary (may not exist yet)
2. ~/.claude/learning/lessons/archive/ — scan all archived lessons for:
   • ## Still Fuzzy On sections
   • ## Key Learnings sections
   • status frontmatter (completed / postponed / abandoned)
   • sessions count vs milestone count (was the lesson harder than estimated?)
3. ~/.claude/learning/logs/ — scan recent session logs (last 30 days) for:
   • ## Struggled With sections
   • ## Still Fuzzy On sections
   • ## Key Learnings sections

What to Produce

Return a learner context to the calling skill (kept in conversation, not written to disk):

fuzzy_topics: Themes that appear in "Still Fuzzy On" across archived lessons or recent logs. Include the source lesson and how recently it was seen. Example:

• "async error handling" — from LangChain lesson, 2 sessions ago
• "state persistence" — from LangGraph lesson, last week

difficulty_adjustment: Per-gap difficulty level based on performance patterns:

• beginner — struggled with basics, needed many hints
• intermediate — completed milestones with some help, some fuzzy areas remain
• advanced — nailed milestones consistently, few or no struggles

Derive this from: nailed-it vs struggle ratio, sessions-taken vs sessions-estimated, and whether fuzzy topics persist or get resolved across lessons.

cross_topic_patterns: Themes that appear as struggles across two or more different lesson topics. Example: "async patterns" struggled in both LangChain and LangGraph lessons.

Graceful Handling

• If insights.md does not exist: skip it, derive what you can from archive and logs
• If archive and logs are empty (new user): return empty context, no errors
• If no patterns are detected: return empty context — don't manufacture insights

Mode 2: Update Insights

Called by /end-learn after the session log and lesson state have been written.

Input

The session's observations from the current conversation:

• Struggles (from the debrief's "Struggled with" section)
• Key learnings (from "Key learnings" section)
• Still fuzzy topics (from "Still fuzzy on" section)
• Hints given during the session

Process

1. Read ~/.claude/learning/insights.md (create it if it doesn't exist)
2. For each struggle theme from this session:
   • If it already exists in ## Recurring Struggles: increment session count, update last-seen date
   • If it's new: add it with count 1
3. For each key learning from this session:
   • If the theme exists in ## Recurring Struggles AND the user demonstrated clear understanding (nailed it, no hints needed): move it to ## Confirmed Strengths
   • Otherwise: leave struggles as-is (one good session doesn't confirm mastery)
4. Update ## Difficulty Trend:
   • Look at the gap addressed by the current lesson
   • If the session had mostly nailed-it observations: trend toward advanced
   • If the session had significant struggles: trend toward beginner
   • Use the overall pattern across sessions, not just this one session
5. If any incidental code observations were noted during the session, add them to ## Incidental Notes with the lesson topic and date
6. Update last_updated in frontmatter

User Notification

After updating, show the user a one-liner ONLY if something meaningful changed:

• A theme moved to confirmed strength: "Insight updated: [theme] moved to confirmed strength"
• A new recurring pattern detected (3+ sessions): "Pattern detected: [theme] has come up in N sessions now"
• Difficulty level changed for a gap: "Difficulty updated: [gap] progressed to [level]"

If nothing meaningful changed (just incremented a count, added a new single-occurrence struggle), say nothing — don't create noise.

Important

• This skill is NEVER called directly by users — only by /learn and /end-learn
• Build Context mode is READ-ONLY — it produces conversation context, writes nothing
• Update Insights mode writes ONLY to insights.md — no other files
• Do not read the repo — repo analysis is handled by /learn step 3.5
• Be conservative with "confirmed strength" promotion — require evidence across multiple sessions, not just one good moment
• Do not manufacture patterns from thin data — if there's only one session logged, there are no cross-topic patterns to report