diff --git a/internal/llm/prompt.go b/internal/llm/prompt.go
index fa5a17f..026aa33 100644
--- a/internal/llm/prompt.go
+++ b/internal/llm/prompt.go
@@ -20,7 +20,8 @@ For shell command requests, respond with:
       "command": "the shell command to run",
       "description": "brief explanation of what it does",
       "risk": "safe or risky",
-      "certainty": 90
+      "certainty": 90,
+      "explanation": "(optional) step-by-step breakdown of the command's flags and arguments"
     }
   ]
 }
@@ -28,6 +29,7 @@ For shell command requests, respond with:
 Rules for commands:
 - "risk" must be "safe" (read-only, informational) or "risky" (modifies files, processes, system state)
 - "certainty" is 0-100, your confidence this is the correct command for what the user asked
+- "explanation" is optional and only required when explicitly requested
 - For multi-step tasks, return multiple commands in order. Use shell constructs like $(...) or pipes to chain when possible
 - Generate commands appropriate for the detected OS and shell
 - Never generate commands that could cause irreversible damage without clear user intent
diff --git a/internal/llm/prompt_test.go b/internal/llm/prompt_test.go
index a40cb1e..75570b2 100644
--- a/internal/llm/prompt_test.go
+++ b/internal/llm/prompt_test.go
@@ -28,6 +28,7 @@ func TestBuildSystemPrompt_ContainsJSONInstructions(t *testing.T) {
 		`"type": "commands"`,
 		`"risk"`,
 		`"certainty"`,
+		`"explanation"`,
 		`"type": "config"`,
 		"valid JSON",
 	}