I want an auto-label “Newsletter_to_do” to include only relevant, original incoming emails from external sources, such as press releases, press invitations, event announcements from members, and job, internship, or freelance opportunities.
The label keeps being applied to emails that should be excluded, especially replies, forwarded messages, and threads involving our organization.
*I want to exclude all replies and follow-ups in conversations involving my domain @domainename.fr.
I’ve tried making these exclusions explicit in the prompt, but the filter still labels all those messages. Is there a recommended way to write this rule, or a limitation in how AI labels evaluate email threads, sender/recipient domains, or the first message in a conversation?
Here are some screenshots to help understand the problem:
The reason this prompt is failing comes down to how Superhuman evaluates conversations and how LLMs process negative conditions inside conversational threads.
Why the Current Rule Fails
Thread-Level Context Leakage: Superhuman’s Auto Label evaluates the entire conversation thread, not just individual incoming messages. If an initial external email meets the criteria for a press release or announcement, the model matches the positive signal. When you reply or forward it, the original positive text is still sitting in the thread history.
Negative Condition Drift: Natural language instructions like “Exclusion prioritaire : si le … a envoyé un mail…” require the LLM to reason across multiple roles, timestamps, and message origins. Language models frequently suffer from “negative constraint neglect”—they latch onto the strong positive topical signals (e.g., communiqué de presse, invitation, offres d’emploi) and overlook the exclusion condition.
Drafting at Character Limits: The prompt in the screenshot is at 993 / 1000 characters, forcing complex French phrasing into a cramped space where attention to exclusions degrades.
The Architecture Fix
Rather than asking an LLM to do deterministic routing that email headers handle much better, split the logic into two parts: deterministic boundaries and semantic classification.
1. Move Hard Exclusions to the UI Filter
Directly below the prompt box, click “Add exclusions”:
Set a hard deterministic rule: From / Sender does not contain @domainename.fr (or use the deterministic filter blocks if available in your rule builder).
If you only want original external emails, use the deterministic exclusion chips to strip your internal domain entirely before the LLM even evaluates the text.
2. Reframe the AI Prompt (Positive Boundaries + Strict Negatives)
When structuring rules for an LLM classifier, avoid narrative conversational French. Use crisp, hierarchical Markdown with explicit positive targets and clear exclusion gates:
Markdown
## Goal
Label only incoming, first-contact announcements from external senders suitable for newsletter publication.
## Must Include
- Original press releases (communiqués de presse)
- Press invitations & member event announcements
- Job, internship, or freelance opportunities (offres d'emploi, stages, piges)
## Strictly Exclude (DO NOT LABEL)
- Any conversation containing a reply, forward, or message from *@domainename.fr*
- Subject lines containing "Re:", "Fwd:", or "Tr:"
- Internal operational discussions, administrative confirmations, or follow-up threads
- Routine correspondence or 1-on-1 member discussions
Best Practices to Lock It Down
Subject Line Signals: Explicitly adding Subject lines starting with Re:, Fwd:, Tr: gives the model a clear syntactic anchor rather than forcing it to analyze conversational chronology.
Active Tuning: Use the / buttons on the right-hand preview panel for the false matches (such as the Re: Communiqué threads currently showing). Superhuman uses those binary signals to calibrate the threshold and reinforce the exclusion rule across the thread evaluation layer.
Edit : it seemed to work better with last settings but you can’t put “@yourdomainname.com” as it’s not a valid request. Be sure to use “*@yourdomainname.com” if you ever want to do this.