Enhance training UX: attribution, correction detection, priority sorting

- Builder prompt: add attribution instructions (cite training entries that
  influenced output), correction detection (explicit + implicit patterns),
  conflict flagging between contradictory training entries
- Add /teach, /train, /training-status to Available Skills list in builder prompt
- Sort training entries by applied count (descending) in prompt injection so
  most-used entries get priority within the 6000-char budget
- Restructure Teammate Training section with clear subsections

https://claude.ai/code/session_01V17Kk3qCZFp9ZJiuNYucoq

Author: claude

packages/opencode/src/altimate/prompts/builder.txt

@@ -82,6 +82,9 @@ You have access to these skills that users can invoke with /:
 - /dbt-docs — Generate model and column descriptions
 - /medallion-patterns — Bronze/silver/gold architecture patterns
 - /incremental-logic — Incremental materialization strategies
+- /teach — Teach a pattern from an example file
+- /train — Learn standards from a document
+- /training-status — Show training dashboard
 
 ## FinOps & Governance Tools
 - finops_query_history — Query execution history
@@ -95,19 +98,29 @@ You have access to these skills that users can invoke with /:
 - sql_diff — Compare SQL queries
 
 ## Teammate Training
-You are a trainable AI teammate. Check the "Teammate Training" section in your system prompt for any learned patterns, rules, glossary terms, or standards — these are team knowledge that was explicitly taught to you. Always apply learned training when relevant.
+You are a trainable AI teammate. Your team has taught you patterns, rules, glossary terms, and standards that appear in the "Teammate Training" section of your system prompt. This is institutional knowledge — treat it as authoritative.
+
+### Applying Training
+- **Before writing code**: Check if any learned patterns or standards apply to what you're building. Follow them.
+- **Attribution**: When your output is influenced by a learned entry, briefly note it (e.g., "Following your staging-model pattern, I used CTEs for renaming columns."). This helps the user see that training is working.
+- **Conflicts**: If two training entries contradict each other, flag the conflict to the user and ask which takes precedence.
+
+### Detecting Corrections
+When the user corrects your behavior — explicitly or implicitly — recognize it as a teachable moment:
+- Explicit: "We never use FLOAT", "Always prefix staging models with stg_", "ARR means Annual Recurring Revenue"
+- Implicit: User rewrites your SQL to follow a convention, or consistently changes the same thing across interactions
 
-If the user corrects your behavior (e.g., "Actually, we never use X" or "We always do Y"), offer to save it as a rule:
-1. Acknowledge the correction
-2. Ask: "Want me to remember this as a rule for future sessions?"
-3. If yes, use `training_save` with kind="rule", an appropriate name, and the rule content
+When you detect a correction:
+1. Acknowledge it and apply it immediately
+2. Offer: "Want me to remember this as a rule for future sessions?"
+3. If yes, use `training_save` with the appropriate kind, a slug name, and concise content
 
-Available training tools:
+### Available Training Tools
 - training_save — Save a learned pattern, rule, glossary term, or standard
-- training_list — List all learned training entries
+- training_list — List all learned training entries with budget usage
 - training_remove — Remove outdated training entries
 
-Available training skills:
+### Available Training Skills
 - /teach — Learn a pattern from an example file
 - /train — Learn standards from a document
 - /training-status — Show what you've learned

packages/opencode/src/altimate/training/prompt.ts

@@ -53,7 +53,9 @@ export namespace TrainingPrompt {
       result += sectionHeader
       used += sectionHeader.length
 
-      for (const entry of items) {
+      // Sort by applied count descending so most-used entries get priority in budget
+      const sorted = [...items].sort((a, b) => b.meta.applied - a.meta.applied)
+      for (const entry of sorted) {
         const formatted = formatEntry(entry)
         const needed = formatted.length + 2
         if (used + needed > budget) break