Marketing automation for developers usually fails in one of two directions. Either nothing is automated and the habit dies within a month, or everything is automated and the output is obviously machine written.
The line between them is not about capability. It is about which part of the job actually benefits.
Automate: remembering
The most common reason developers do not post is not that they lack material. It is that by Friday they cannot remember Tuesday.
This is a pure retrieval problem and the record already exists. Automating it costs you nothing and solves the most common failure.
Automate: filtering
A working week produces a lot of commits and most of them are noise. Merge commits, dependency bumps, formatting, version tags.
Deciding which are noise is mechanical and consistent. There is no judgment being lost here.
Automate: grouping
"Added export button", "Added CSV export", "Fixed empty export case" are one piece of work. Recognising that is pattern matching over scopes, file paths and vocabulary, which a machine does about as well as you would and much faster.
Automate: first drafts
Turning structured facts into readable prose is genuinely what language models are good at, and a blank page is where most posting attempts die.
The important constraint is that the draft must be built from evidence rather than from a prompt. A draft generated from four real commits is a starting point. A draft generated from "write about shipping a feature" is a liability.
Never automate: publishing
This is the line.
Once something goes out under your name without you reading it, every post you have ever made becomes slightly less trustworthy, because readers now know that some of them were not read by you.
The saving is small. Review takes thirty seconds. The risk is unbounded: one wrong claim, one leaked internal detail, one tone-deaf post on a bad day.
Never automate: the point
A model can describe what changed. It cannot know why it mattered, because the why lives in conversations, constraints and decisions that are not in the code.
The opinion is the part people read for. Automating it produces content that is technically accurate and entirely forgettable.
Never automate: replies
Someone takes ten minutes to write a thoughtful comment and gets a generated reply. If they notice, you have lost them permanently, and people notice more often than you would expect.
What the workflow ends up being
Automated: scan, filter, group, score, draft.
Manual: read, correct, add the one sentence only you could write, publish.
About thirty seconds of human time per post, applied at the only point where human time is worth anything.
That is the whole argument. Automate the recall and the typing. Keep the judgment.