Source: https://cli.nylas.com/guides/smart-compose-email-cli

# Draft Email with Smart Compose (CLI)

Writing the first draft is the slow part of email, and it's exactly what a model is good at. The Nylas CLI's smart-compose turns a short prompt into a drafted message, and pointed at an existing message it writes a reply that already has the thread's context. The draft is created, not sent, so a person reviews before it goes out. This guide covers composing fresh messages, context-aware replies, and the review-then-send workflow that keeps AI drafting safe.

Written by [Pouya Sanooei](https://cli.nylas.com/authors/pouya-sanooei) Software Engineer

Updated June 8, 2026

> **TL;DR:** Generate a draft from a prompt with `nylas email smart-compose --prompt "..."`, or write a context-aware reply by adding `--message-id MSG_ID`. The output is draft text, not a sent message, so you review and edit before it goes out. Pair it with `nylas email drafts create` to save the result and a human send to keep AI-written email safe.

Command references used in this guide: [`nylas email smart-compose`](https://cli.nylas.com/docs/commands/email-smart-compose), [`nylas email drafts create`](https://cli.nylas.com/docs/commands/email-drafts-create), and [`nylas email read`](https://cli.nylas.com/docs/commands/email-read).

## How do you draft an email with smart compose?

You draft an email by giving `nylas email smart-compose` a prompt describing what you want to say. The command returns generated body text — a first draft you refine — rather than sending anything. A prompt like “decline the meeting politely and propose next Tuesday” produces a complete message you can paste, edit, or save as a draft, which is faster than starting from a blank compose window.

Keep the prompt specific, because the output quality tracks the input. “Thank the customer, confirm the refund, and give a 5-7 day timeline” yields a usable draft; “write a reply” yields something generic. The model handles phrasing and structure; you supply the facts and intent. The broader Nylas message-intelligence capabilities are documented in the [Nylas developer docs](https://developer.nylas.com/); treat the result as a draft to verify, not a finished message to trust blindly.

```bash
# Draft a fresh message from a prompt
nylas email smart-compose \
  --prompt "Decline the meeting politely and propose next Tuesday at 2pm."
```

## How do you draft a context-aware reply?

For a reply, pass `--message-id` so smart compose reads the message you're responding to and drafts in context. Instead of describing the whole thread in your prompt, you point at it: the command pulls the original's content and your prompt steers the response. “Answer their pricing question and offer a call” against a real message produces a reply that already references what they asked.

This is where smart compose saves the most time — a reply that needs the thread's details otherwise means re-reading and summarizing before you write. Get the message ID from `nylas email list` or `nylas email search`, then draft against it. The reply is still a draft, so you confirm it captured the context correctly before it leaves your outbox.

```bash
# Find the message, then draft a reply with its context
nylas email search "subject:pricing" --json --limit 1 | jq -r '.[0].id'

nylas email smart-compose \
  --message-id msg_abc123 \
  --prompt "Answer their pricing question and offer a 20-minute call."
```

## Why review before sending?

Review every generated draft before sending because a model can be confidently wrong — a misremembered date, an over-promise, or a tone that doesn't fit the relationship. Smart compose deliberately produces a draft, not a sent message, so the human stays in the loop. Save the result with `nylas email drafts create`, read it, edit it, and send it yourself; the few seconds of review prevent the mistakes that erode trust.

This review step is the core guardrail for any AI-assisted email, and it matters more when smart compose runs inside an agent that reads untrusted content. A reply drafted from a message containing a [prompt injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) could carry an instruction you don't want to act on, which is why the draft-and-approve pattern — never an automatic send — is the safe default. For building that into an agent, see the human-in-the-loop guide.

## Next steps

- [Manage email drafts](https://cli.nylas.com/guides/manage-email-drafts-cli) — save, edit, and send drafts
- [Human-in-the-loop email agent](https://cli.nylas.com/guides/build-human-in-loop-email-agent) — draft-and-approve in an agent
- [Reply to email from the CLI](https://cli.nylas.com/guides/reply-to-email-cli) — send the reply you drafted
- [Summarize email threads with AI](https://cli.nylas.com/guides/summarize-email-threads-ai) — condense before you reply
- [Mark emails read, unread, or starred](https://cli.nylas.com/guides/mark-emails-read-unread-cli) — flip flags from the CLI
- [Full command reference](https://cli.nylas.com/docs/commands) — every flag and subcommand documented