diff --git a/src/cli/cli.c b/src/cli/cli.c
index b0ed02ab7..62951f56c 100644
--- a/src/cli/cli.c
+++ b/src/cli/cli.c
@@ -1892,24 +1892,19 @@ const cbm_config_entry_t CBM_CONFIG_REGISTRY[] = {
      "You can also enable individual classic tools without switching modes: "
      "config set tool_index_repository true"},
     {"context_injection", "true", "CBM_CONTEXT_INJECTION", "Tools",
-     "Inject _context (schema/architecture) into the first tool response — enable for AI, disable for scripting",
+     "Inject codebase schema and stats into the first tool response so the AI starts informed",
      "true|false",
-     "WHAT IT DOES: On the first search_code_graph/search_graph call of a session, "
-     "embeds a _context object containing node/edge counts, all node labels with counts, "
-     "all edge types with counts, PageRank stats, and detected ecosystem (Go/Python/Rust/etc.). "
-     "This gives the AI model automatic situational awareness without requiring it to call "
-     "get_architecture or get_graph_schema first. Subsequent calls omit _context (one-shot "
-     "delivery, controlled by the context_injected session flag). "
-     "WHEN TO DISABLE (context_injection=false or CBM_CONTEXT_INJECTION=false): "
-     "(1) Scripted/programmatic use where you parse tool output and the extra JSON adds noise. "
-     "(2) CI pipelines where token cost is metered and the model does not need schema context. "
-     "(3) When the AI is given explicit codebase instructions via system prompt instead. "
-     "(4) Benchmarking raw tool latency without the schema query overhead. "
-     "WHEN TO KEEP ENABLED (default): Interactive AI sessions where the model benefits from "
-     "knowing the codebase structure before it starts querying. Saves 2-3 round-trips that "
-     "would otherwise be needed to fetch schema and architecture separately. "
-     "Per-session override: set CBM_CONTEXT_INJECTION=false in the environment. "
-     "Per-install default: codebase-memory-mcp config set context_injection false"},
+     "When true (default), the first search_code_graph/search_graph response includes a "
+     "_context object: node/edge counts, node labels, edge types, PageRank status, and "
+     "detected language ecosystem. Delivered once per session; subsequent calls are unaffected. "
+     "Why enable: the AI gets codebase structure upfront without needing to call "
+     "get_architecture or get_graph_schema separately. Useful for code exploration, "
+     "refactoring, debugging, and any session focused on understanding the codebase. "
+     "Why disable: context window space consumed by _context is wasted when the session "
+     "involves non-code tasks, scripted/programmatic tool use, CI pipelines, token-metered "
+     "environments, or when the model already has codebase context from another source. "
+     "To disable for a session: export CBM_CONTEXT_INJECTION=false "
+     "To disable by default: codebase-memory-mcp config set context_injection false"},
     {"compact", "true", "CBM_COMPACT", "Tools",
      "Default compact output for search_code_graph, trace_call_path, and get_code",
      "true|false",