You're doing outreach in dev communities. The first ten replies got real responses. By the thirtieth, replies have dried up and a couple of people called your account out as a bot. You hadn't changed the content. You'd changed nothing, in fact — that was the problem.

The thing that gives you away isn't the words. It's the rhythm.

What "rhythm" means in writing

Every reader has a fast-running, mostly-unconscious model of "how a real comment sounds." Sentence lengths vary. Sentences sometimes start with And or Tbh. The opener isn't always the same shape. Conclusions aren't always present.

When you author the same kind of reply 30 times — even with the actual sentences swapped out — your macro structure is the constant. Opener with a hook, three sentences of advice, a soft CTA. That macro shape becomes the signal. People don't consciously detect it, but their bot-detector fires anyway.

It's the same reason a song with the same beat under different melodies still sounds like one band's sound.

The four rhythm dimensions that leak templates

When you audit your last 20 outreach replies side by side, look at:

Opener shape. Are they all questions? All "This sounds like…"? All "Have you tried…"? Pick three different openers and rotate them.
Sentence-length distribution. Templates tend to medium sentences. Real replies have one-word sentences. Long, comma-spliced ones. Fragments. Mix it up.
Paragraph count. Three short paragraphs every time is a template smell. Sometimes one paragraph is the right answer. Sometimes seven.
Signoff. "Hope that helps!" — at the bottom of every reply — is a tell so loud you might as well sign "Sent from my Marketing Department."

Vary all four and you can repeat the same content angle without anyone noticing.

How to audit yourself

Open your last 20 outreach replies in one document. Scroll through them quickly without reading the words. Just look at the shape — paragraph count, length, indentation. If they look identical when blurred, you have a template problem regardless of what the actual sentences say.

A second pass: read just the first sentence of each reply, in order, as a list. If they sound like a poem with a meter, that's bad. If they sound like 20 different humans, that's good.

A worked rewrite

Three of your replies, in their current form:

Sounds like a connection pool issue. Have you checked pg_stat_activity? Often it's a long-running transaction. Hope that helps!

Sounds like a TLS handshake mismatch. Have you checked the SNI config? Often it's a wildcard cert issue. Hope that helps!

Sounds like a stale Docker layer. Have you checked the build cache? Often it's a hidden COPY. Hope that helps!

Same skeleton three times. The reader's pattern detector wakes up.

Same content, varied rhythm:

Connection pool — pg_stat_activity should show idle-in-transaction rows if that's the cause. Bet on a transaction not being closed in an error path.

Have you tried disabling SNI verification temporarily and seeing if the handshake completes? If yes, your wildcard cert isn't matching the host the client thinks it's connecting to.

Build cache. Specifically a stale layer holding an old COPY. docker build --no-cache once will tell you. If the new build works, you have a layer-invalidation gap somewhere upstream.

Three different shapes, same density of information.

Why this is worth caring about

You can write objectively correct, helpful content and still get ignored — or worse, downranked — if the rhythm signals that you're a bot doing rounds. Once a community labels you that way, every subsequent reply gets read with that prior. Reversing it costs more time than varying the rhythm cost in the first place.

The tactical version is small: vary your openers, sentence lengths, and signoffs. The strategic version is bigger: stop thinking of your outreach as a campaign with a template, and start thinking of each reply as one specific person responding to one specific question.