feat(ce-plan): reduce token usage by extracting conditional references

plugins/compound-engineering/AGENTS.md

@@ -93,6 +93,10 @@ When adding or modifying skills, verify compliance with the skill spec:
   This resolves relative to the SKILL.md and substitutes content before the model sees it. If a file is over ~150 lines, prefer a backtick path even if it is always needed
 - [ ] For files the agent needs to *execute* (scripts, shell templates), always use backtick paths -- `@` would inline the script as text content instead of keeping it as an executable file
 
+### Conditional and Late-Sequence Extraction
+
+Skill content loaded at trigger time is carried in every subsequent message — every tool call, agent dispatch, and response. This carrying cost compounds across the session. For skills that orchestrate many tool or agent calls, extract blocks to `references/` when they are conditional (only execute under specific conditions) or late-sequence (only needed after many prior calls) and represent a meaningful share of the skill (~20%+). The more tool/agent calls a skill makes, the more aggressively to extract. Replace extracted blocks with a 1-3 line stub stating the condition and a backtick path reference (e.g., "Read `references/deepening-workflow.md`"). Never use `@` for extracted blocks — it inlines content at load time, defeating the extraction.
+
 ### Writing Style
 
 - [ ] Use imperative/infinitive form (verb-first instructions)