Composing Emails

Composing emails

The block model

Emit content as an array of { tag, content } blocks. Only three tags exist:

• paragraph → <p> body copy
• heading → <h2> section header
• link → <a> for URLs

No lists, tables, images, bold/italic, blockquotes, or code blocks. If you need emphasis, use a heading + paragraph. If you need a list, write each item as its own paragraph. All strings are HTML-escaped, so don't try to inject tags.

Formatting rules

• No markdown. No # headings, no **bold**, no - bullets — they render literally.
• No newlines inside a block's content — break into separate blocks instead.
• Preserve URLs from source data as link blocks, not inline in paragraph text.
• Format numbers, prices, and dates clearly inside paragraph copy (e.g. $1,234.56, April 28, 2026).

Tone

• Professional but friendly
• Descriptive subject line — not "Update" or "FYI"
• Lead with the point; don't bury it under context
• One idea per paragraph

Recipient sourcing

Always read the recipient address from a typed signal-payload field interpolated into the action prompt — never from upstream agent output. LLM agents will confabulate plausible-looking addresses when the field is absent rather than failing.

Wire it end-to-end:

1. Signal schema declares the recipient as a required field — caller can't trigger the job without it.

   signals:
     send-alert:
       schema:
         type: object
         properties:
           notify_email: { type: string, format: email }
         required: [notify_email]
   

2. Action prompt interpolates from the signal payload (works in any step, including ones with inputFrom — see writing-workspace-jobs → "Signal payload threading"):

   - type: agent
     agentId: emailer
     prompt: "Send the alert to {{inputs.notify_email}}"
   

3. Email-sender agent system prompt should refuse rather than guess: "If the recipient address contains an unresolved {{...}} placeholder or is missing, call failStep with a reason. Never invent a recipient."

The literal-placeholder check is the important part — the FSM engine intentionally leaves unresolved placeholders verbatim so authoring bugs stay visible. A guarded agent treats that as a hard stop.