diff --git a/src/azure/chatgpt.rs b/src/azure/chatgpt.rs
index c67ac36..716638c 100644
--- a/src/azure/chatgpt.rs
+++ b/src/azure/chatgpt.rs
@@ -34,6 +34,7 @@ pub struct ChatGPT {
     messages: Rc<Vec<ChatRequestMessage>>,
     tools: Option<Rc<[Tool]>>,
     function_store: FunctionStore,
+    last_assistant_message: String,
     pub listener: Option<Box<dyn ChatListener>>,
 }
 
@@ -58,11 +59,12 @@ impl ChatGPT {
             messages: Rc::new(vec![]),
             tools,
             function_store,
+            last_assistant_message: String::new(),
             listener: None,
         }
     }
 
-    pub async fn chat(&mut self, message: String, files: Option<Vec<PathBuf>>) -> Result<(), Exception> {
+    pub async fn chat(&mut self, message: String, files: Option<Vec<PathBuf>>) -> Result<String, Exception> {
         let image_urls = image_urls(files).await?;
         self.add_message(ChatRequestMessage::new_user_message(message, image_urls));
 
@@ -82,7 +84,7 @@ impl ChatGPT {
             }
             self.process().await?;
         }
-        Ok(())
+        Ok(self.last_assistant_message.to_string())
     }
 
     pub fn system_message(&mut self, message: String) {
@@ -143,6 +145,7 @@ impl ChatGPT {
         }
 
         if !assistant_message.is_empty() {
+            self.last_assistant_message = assistant_message.to_string();
             self.add_message(ChatRequestMessage::new_message(Role::Assistant, assistant_message));
         }
 
diff --git a/src/command.rs b/src/command.rs
index 1fdbb17..7c48dfe 100644
--- a/src/command.rs
+++ b/src/command.rs
@@ -1,3 +1,4 @@
 pub mod chat;
+pub mod complete;
 pub mod generate_zsh_completion;
 pub mod speak;
diff --git a/src/command/complete.rs b/src/command/complete.rs
new file mode 100644
index 0000000..849a65f
--- /dev/null
+++ b/src/command/complete.rs
@@ -0,0 +1,101 @@
+use std::io::Write;
+use std::path::PathBuf;
+
+use clap::Args;
+use tokio::fs;
+use tokio::io::AsyncBufReadExt;
+use tokio::io::AsyncWriteExt;
+use tokio::io::BufReader;
+use tracing::info;
+
+use crate::llm;
+use crate::llm::ChatEvent;
+use crate::llm::ChatListener;
+use crate::util::exception::Exception;
+
+#[derive(Args)]
+pub struct Complete {
+    #[arg(help = "prompt file")]
+    prompt: PathBuf,
+
+    #[arg(long, help = "conf path")]
+    conf: PathBuf,
+
+    #[arg(long, help = "model name")]
+    name: String,
+}
+
+struct Listener;
+
+impl ChatListener for Listener {
+    fn on_event(&self, event: ChatEvent) {
+        match event {
+            ChatEvent::Delta(data) => {
+                print!("{data}");
+                let _ = std::io::stdout().flush();
+            }
+            ChatEvent::End(usage) => {
+                println!();
+                info!(
+                    "usage, request_tokens={}, response_tokens={}",
+                    usage.request_tokens, usage.response_tokens
+                );
+            }
+        }
+    }
+}
+
+impl Complete {
+    pub async fn execute(&self) -> Result<(), Exception> {
+        let config = llm::load(&self.conf).await?;
+        let mut model = config.create(&self.name)?;
+        model.listener(Box::new(Listener));
+
+        let prompt = fs::OpenOptions::new().read(true).open(&self.prompt).await?;
+        let reader = BufReader::new(prompt);
+        let mut lines = reader.lines();
+
+        let mut files: Vec<PathBuf> = vec![];
+        let mut message = String::new();
+        let mut on_system_message = false;
+        loop {
+            let Some(line) = lines.next_line().await? else { break };
+
+            if line.is_empty() {
+                continue;
+            }
+
+            if line.starts_with("# system") {
+                if !message.is_empty() {
+                    return Err(Exception::ValidationError("system message must be at first".to_string()));
+                }
+                on_system_message = true;
+            } else if line.starts_with("---") || line.starts_with("# file: ") {
+                if on_system_message {
+                    info!("system message: {}", message);
+                    model.system_message(message);
+                    message = String::new();
+                    on_system_message = false;
+                }
+                if line.starts_with("# file: ") {
+                    let file = PathBuf::from(line.strip_prefix("# file: ").unwrap().to_string());
+                    info!("file: {}", file.to_string_lossy());
+                    files.push(file);
+                }
+            } else {
+                message.push_str(&line);
+                message.push('\n');
+            }
+        }
+
+        info!("prompt: {}", message);
+        let files = files.into_iter().map(Some).collect();
+        let message = model.chat(message, files).await?;
+
+        let mut prompt = fs::OpenOptions::new().append(true).open(&self.prompt).await?;
+        prompt.write_all(b"\n---\n\n").await?;
+        prompt.write_all(message.as_bytes()).await?;
+
+        Ok(())
+    }
+}
diff --git a/src/gcloud/gemini.rs b/src/gcloud/gemini.rs
index 5c7289f..547fb22 100644
--- a/src/gcloud/gemini.rs
+++ b/src/gcloud/gemini.rs
@@ -34,10 +34,11 @@ use crate::util::json;
 pub struct Gemini {
     url: String,
     messages: Rc<Vec<Content>>,
-    system_message: Option<Rc<Content>>,
+    system_instruction: Option<Rc<Content>>,
     tools: Option<Rc<[Tool]>>,
     function_store: FunctionStore,
     usage: Usage,
+    last_model_message: String,
     pub listener: Option<Box<dyn ChatListener>>,
 }
 
@@ -47,17 +48,18 @@ impl Gemini {
         Gemini {
             url,
             messages: Rc::new(vec![]),
-            system_message: None,
+            system_instruction: None,
             tools: function_store.declarations.is_empty().not().then_some(Rc::from(vec![Tool {
                 function_declarations: function_store.declarations.to_vec(),
             }])),
             function_store,
             usage: Usage::default(),
+            last_model_message: String::with_capacity(1024),
             listener: None,
         }
     }
 
-    pub async fn chat(&mut self, message: String, files: Option<Vec<PathBuf>>) -> Result<(), Exception> {
+    pub async fn chat(&mut self, message: String, files: Option<Vec<PathBuf>>) -> Result<String, Exception> {
         let data = inline_datas(files).await?;
         self.add_message(Content::new_user_text(message, data));
 
@@ -67,11 +69,11 @@ impl Gemini {
             self.add_message(Content::new_function_response(function_call.name, function_response));
             result = self.process().await?;
         }
-        Ok(())
+        Ok(self.last_model_message.to_string())
     }
 
-    pub fn system_message(&mut self, message: String) {
-        self.system_message = Some(Rc::new(Content::new_model_text(message)));
+    pub fn system_instruction(&mut self, message: String) {
+        self.system_instruction = Some(Rc::new(Content::new_model_text(message)));
     }
 
     async fn process(&mut self) -> Result<Option<FunctionCall>, Exception> {
@@ -126,6 +128,7 @@ impl Gemini {
         }
 
         if !model_message.is_empty() {
+            self.last_model_message = model_message.to_string();
             self.add_message(Content::new_model_text(model_message));
         }
 
@@ -144,7 +147,7 @@ impl Gemini {
     async fn call_api(&self) -> Result<Response, Exception> {
         let request = StreamGenerateContent {
             contents: Rc::clone(&self.messages),
-            system_instruction: self.system_message.clone(),
+            system_instruction: self.system_instruction.clone(),
             generation_config: GenerationConfig {
                 temperature: 1.0,
                 top_p: 0.95,
diff --git a/src/llm.rs b/src/llm.rs
index 2b16c25..e01804f 100644
--- a/src/llm.rs
+++ b/src/llm.rs
@@ -34,7 +34,7 @@ pub enum Model {
 }
 
 impl Model {
-    pub async fn chat(&mut self, message: String, files: Option<Vec<PathBuf>>) -> Result<(), Exception> {
+    pub async fn chat(&mut self, message: String, files: Option<Vec<PathBuf>>) -> Result<String, Exception> {
         match self {
             Model::ChatGPT(model) => model.chat(message, files).await,
             Model::Gemini(model) => model.chat(message, files).await,
@@ -51,7 +51,7 @@ impl Model {
     pub fn system_message(&mut self, message: String) {
         match self {
             Model::ChatGPT(model) => model.system_message(message),
-            Model::Gemini(model) => model.system_message(message),
+            Model::Gemini(model) => model.system_instruction(message),
         }
     }
 }
diff --git a/src/main.rs b/src/main.rs
index d3e44ba..5bdf291 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -1,6 +1,7 @@
 use clap::Parser;
 use clap::Subcommand;
 use command::chat::Chat;
+use command::complete::Complete;
 use command::generate_zsh_completion::GenerateZshCompletion;
 use command::speak::Speak;
 use util::exception::Exception;
@@ -28,6 +29,8 @@ pub enum Command {
     Chat(Chat),
     #[command(about = "speak")]
     Speech(Speak),
+    #[command(about = "complete")]
+    Complete(Complete),
     #[command(about = "generate zsh completion")]
     GenerateZshCompletion(GenerateZshCompletion),
 }
@@ -39,6 +42,7 @@ async fn main() -> Result<(), Exception> {
     match cli.command {
         Some(Command::Chat(command)) => command.execute().await,
         Some(Command::Speech(command)) => command.execute().await,
+        Some(Command::Complete(command)) => command.execute().await,
         Some(Command::GenerateZshCompletion(command)) => command.execute(),
         None => panic!("not implemented"),
     }