If your product has a public email address that does something, you will eventually learn the same lesson I did: the address gets found by people who want to sell you things. Lead lists, "quick question about your growth", offers to upvote your launch. They arrive at the same address as real work.

I build Maildoer, where you email a task and a file to an address and get the result back by email. So for me, a pitch isn't just noise. If it gets treated as a task, the system does work nobody asked for, writes a reply nobody wants, and spends the sender's trial credits on a sales email. This post is the decision rule I ended up with. It applies to any shared inbox where real requests and pitches mix: a support address, a careers address, an "info@".

What the mix actually looked like

At one point I lined up a stretch of recent mail and counted: 9 pitches against 35 real tasks. Small numbers, but the pitches were not obviously different at a glance. Several were long, polite, and well formatted. A few even attached a PDF. My first, too-lenient rules let some of them through as tasks. Long emails were assumed to be real, and that was the mistake.

Three buckets, not two

The thing that helped most was admitting there are three kinds of mail, not "real vs spam":

  1. Real requests. Someone wants the thing your address does.
  2. Pitches. Someone wants you to buy, reply, or promote something.
  3. Genuine partnership inquiries. Rare and worth a human reply, but not an automatic one.

Lumping 3 into 2 means you miss the one email that matters. Lumping 3 into 1 means an automated reply goes to a potential partner, which is worse.

The checklist I use to classify by hand

When I review a borderline message myself, I go through these questions in order and stop at the first clear answer:

1. Does it ask the address to DO its job (summarize, translate, analyze, draft)? -> real request 2. Is the main ask "reply / book a call / buy / upvote / feature us"? -> pitch 3. Does it propose working together on terms (reseller, integration, co-marketing)? -> partnership, human reply 4. Would the sender be annoyed to receive the product's normal output? -> not a real request 5. Still unsure? Treat as real, but review before anything is charged or sent.

Question 4 catches the polished ones. A growth agency does not want a ten-slide summary of its own cold email.

What happens to each bucket

  • Real requests go through as normal.
  • Pitches are archived with a note, not processed, not answered, and nothing is charged to anyone.
  • Partnership inquiries go to a separate folder that a human (me) reads. No auto-reply.

The "archived, not deleted" part matters. When I tighten rules, I want to look back and check that no real user got caught.

Where memory changes the picture

The case I worried about most was a real user's second email. Their first message was a clear task with a spreadsheet. Their second one says only: "Same as last week, but add the new store." Read alone, that line has no task verb and no file. It could look like anything.

But it is not alone. In Maildoer, a later email from the same person is read together with their earlier ones, so "same as last week" points back to the earlier spreadsheet and instructions instead of starting from a blank page. That is the whole point of the product, and it is also why I don't want a triage rule that judges every email as if it were the sender's first. For your own inbox, the hand-checklist version is: before you file a short, vague email as noise, check whether that sender has written to you before.

What I'd do differently

  • Start stricter. I began lenient because a false "pitch" on a real user felt worse than a missed pitch. In practice, the pitches that slipped through cost real credits and produced replies nobody read. Strict plus an archive you actually review is safer than lenient.
  • Never use length as a signal. Long emails are not more likely to be real. Mine were often the opposite.
  • Keep the partnership folder small and read it. If it becomes a second spam folder, the rule has failed.
  • Don't refund silently or charge silently. Decide the credit policy for misfiled mail up front and write it down.

The short version

Three buckets. Classify by the ask, not the length. Archive pitches, read partnerships yourself, and before calling a vague email noise, check what that person sent before.

How do you handle pitches on a public address that triggers real work? Do you filter, auto-reply, or just eat the cost?

If you want to see the "same as last week" behaviour yourself: email a file and a task to do@maildoer.com, then send a short follow-up that assumes it remembers. Each address starts with 100 credits. https://maildoer.com