What this prompt does
This prompt asks the AI to build an AI email assistant for a [user_role] using [model_provider]. It classifies incoming mail into [categories] with confidence scores, drafts [reply_variants] reply variants (brief, detailed, formal), and adapts to the user's writing style by learning from [sample_count] sent emails. It extracts action items, deadlines, and follow-ups, detects scheduling requests, scores priority from sender importance and urgency signals, and summarizes threads longer than [thread_threshold] messages.
The structure works because it splits the inbox problem into discrete, testable capabilities rather than one vague "AI for email" feature. Pairing a fast model for classification with a stronger one for drafting (as [model_provider] suggests) keeps cost and latency sane. PII detection flags sensitive information before it leaves your draft, and integration with [email_provider] via API or IMAP, implemented in [language] with background processing, makes it safe enough to point at a real inbox. Priority scoring is the quiet workhorse here: by weighing sender importance, urgency signals, and content together, it surfaces the few emails that actually need a reply now, which is the difference between an assistant that saves time and one that just adds another queue to read.
When to use it
- You spend too long triaging and replying to email and want classification plus draft variants to speed it up.
- You want the assistant to learn your tone from
[sample_count]past emails instead of sounding generic. - You need action items and deadlines pulled out of threads automatically.
- You want priority scoring so urgent and important mail surfaces first.
- You handle sensitive data and need PII flagged before anything is sent.
- You are integrating against
[email_provider]and want background classification rather than blocking the UI.
Example output
Expect a classification module tagging mail into [categories] with confidence, a drafting module producing [reply_variants] tonal variants, an action-item extractor, a thread summarizer triggered past [thread_threshold] messages, a priority scorer, and a PII detector. You also get the drafting system prompt, [email_provider] integration code in [language], and evaluation metrics for classification accuracy.
Pro tips
- Set
[model_provider]as a split — a fast model for[categories]classification, a stronger one for drafting — to balance latency and quality. - Tune
[categories]to your actual workflow; generic buckets produce generic routing, so name the categories you really triage by. - For style learning, more
[sample_count]is not always better — use representative sent emails, since a few off-tone samples can skew the voice. - Treat PII detection as advisory, not a guarantee; flag and warn, but keep a human in the loop before sensitive drafts send.
- Calibrate confidence thresholds so low-confidence classifications route to manual review instead of silently mis-filing mail.
- Start with
[reply_variants]as draft suggestions a human edits, not auto-send, until you trust the tone on your own inbox.
