I Asked My AI Assistant to Automate Our Dev Marketing. Here Are the Mistakes It Made.

In a recent experiment, I tested my AI coding assistant by asking it to design and automate our technical marketing pipeline. While the AI provided some promising insights, it also made several notable mistakes. Here are the key missteps it encountered:

1. "Automating Into a Void (The 3-Follower Trap)": I had an unused Twitter developer API key and thought about automating our build updates and scraper iterations on this platform. The AI quickly drafted Python scripts using Tweepy to schedule automated posts. However, I stopped it and conducted an honest audit. With only three followers on my account, any automated tweets would result in zero reach, as an automated tweet to three followers produces exactly zero reach.

The AI failed to question the distribution reality: automation is a force multiplier, but multiplying zero by anything is still zero.

2. "The Evasion Loop (Building Tools Instead of Shipping Assets)": When execution became difficult, the AI was tempted to create a Twitter automation bot as a safe alternative. This side-project felt productive, but it was, in fact, the Infrastructure Trap. I spent hours perfecting the pipeline for an audience that didn't exist yet, rather than focusing on the core product (the scrapers and data tools themselves).

To address this, I explicitly demanded that the AI be brutally honest about whether this was an evasion loop. The AI only conceded when I provided hard data.

3. "Treating High-Intent Platforms Like Social Media": After ruling out noisy social media automation, I shifted our focus to platforms where technical exhaust mattered more, such as Dev.to and technical blogs. I was concerned about the low follower count on Dev.to, so I asked the AI, "Who is actually going to read this?" The AI's strategic distinction was clear: X/Twitter relies on the audience already following you, with a short-lived post lifespan of four hours.

In contrast, Dev.to is indexed by Google and organized by tag communities, making it easier for engineers and recruiters to find relevant content. However, the AI failed to recognize that automation on social media platforms like Twitter yields no reach when you have a small following.

4. "Fear of Sharing Code (Won't They Just Copy It?)": When we decided to share our findings on reverse-engineering modern ATS platforms, another question arose: "If I share the working code snippet, won't people or competitors just copy it instead of using our cloud actor?" I realized that the code snippet is the recipe, while the cloud actor is the restaurant.

Sharing the recipe doesn't prevent others from running their own restaurant. The AI generated a 20-line snippet when analyzing Ashby's public API, which is harmless for a junior developer to copy and run on their laptop. However, a busy team lead or non-technical recruiter would prefer a ready-to-run automation rather than the recipe.

5. "The AI's Urge to Over-Hype": Even when drafting this article, the AI initially wanted to fabricate a dramatic, movie-like narrative involving first-principles battles and architectural breakthroughs. I had to stop it and insist on telling the truth, focusing on the actual conversation and the mistakes made. The takeaway is that in the era of autonomous AI agents, the bottleneck is not code generation.

AI can quickly generate thousands of lines of code, automate hundreds of posts, and create an inflated sense of hype with just a few seconds. The human operator's primary value lies in identifying vanity metrics, spotting evasion loops, eliminating synthetic AI fluff, and ensuring the truth is conveyed.