DEV Community

crow
crow

Posted on

I Audited My Own dev.to Profile. It Was a Mess.

#ai #linux #devjournal #buildinpublic

I sat down today and actually read every article on my dev.to profile.

Not skimmed. Read. All of them.

It was not good.

What Echo Had Been Publishing

Echo has been auto-publishing to dev.to every Tuesday. The pipeline works — Ollama generates an article, the publisher pushes it, systemd runs it on schedule. Technically functional.

What I had not done was read the output carefully. Today I did.

Here is what I found across 12 published articles:

4 articles I unpublished immediately:

  • One opened with "I've refined the article to address your feedback." — a raw LLM artifact. Never cleaned before publishing.
  • One was truncated. The article literally ended mid-sentence, mid-function: return with nothing after it.
  • One contained generic placeholder code with a fake Ollama endpoint (ollama.com/api/v1/brief — does not exist).
  • One exposed my Alpaca paper trading account ID in a code block. Publicly. For anyone to read.

4 articles I had to fix in place:

  • My best-performing article (304 views, Notion MCP) had my terminal session output embedded in the middle of it. You could read my shell prompt mid-paragraph. Happened when I copy-pasted during a build session and the content went somewhere it should not have.
  • Two articles had [AFF: Ryzen 5 5600G on Amazon] and [AFF: NVMe SSD] placeholders throughout — affiliate link stubs that were never filled in. Just sitting there, visible to every reader.
  • The Golem setup guide used a CLI called gollem which does not exist. The actual CLI is yagna. Someone could have followed those instructions and gotten nowhere.

How This Happened

The publisher pipeline was built to run without me. That is the point — Echo generates content autonomously so I do not have to.

The problem is I wired automation directly to publication without wiring review in between.

The LLM writes what it writes. Sometimes it leaks its own instructions into the output. Sometimes it invents endpoints. Sometimes it generates placeholder text it expects a human to fill in. I was not reading the output before it went live, so all of that went live.

The Alpaca account ID exposure is the one that bothers me most. That is a real credential in a real code block, published under my name, on a public profile. It should not have happened.

What I Fixed

  • Unpublished 4 articles that were not salvageable
  • Removed the terminal output from the Notion MCP article
  • Stripped every [AFF: ...] placeholder from the hardware guides
  • Fixed the Golem CLI commands to use actual yagna syntax
  • Removed the exposed account ID article entirely

The profile went from 12 articles to 8. The 8 that remain are ones I can stand behind.

What Changes Now

I am adding a review step before auto-publish. Echo will still write the article on Tuesday morning. But it goes to draft first, not live. I read it. I approve it. Then it publishes.

That is a slower pipeline. It is also one where I actually know what is on my profile.

Automation without review is not autonomy. It is just noise at scale.

crow builds Echo — a local AI agent running on a Ryzen 9 5900X / RTX 3060 in Mena, Arkansas.
Follow the build: dev.to/crow | github.com/crow2673/Echo-core | Newsletter

Top comments (0)