Community/LinkedIn post about new product
This experiment compares how different LLMs handle a specific marketing task: writing a LinkedIn launch post for a macOS tool called Oshn Prompt Generator. Each model was evaluated based on three key pillars: 1) Publish-Ready Quality: Would I actually post this as-is, or does it require heavy editing? 2) AI Fingerprint: Does the text feel "robotic" and full of AI cliches, or does it sound like a natural human creator? 3) Platform Optimization: How well does the model leverage LinkedIn-specific mechanics, such as effective hook lines, spacing, relevant hashtags, and viral formatting?
28 experiments14 models tested
GPT-5 MiniGPT-5 NanoClaude Haiku 4.5Claude Opus 4.5GPT-5GPT-5.2Claude Sonnet 4.5Gemin…
Community/LinkedIn post about new product
This experiment compares how different LLMs handle a specific marketing task: writing a LinkedIn launch post for a macOS tool called Oshn Prompt Generator. Each model was evaluated based on three key pillars: 1) Publish-Ready Quality: Would I actually post this as-is, or does it require heavy editing? 2) AI Fingerprint: Does the text feel "robotic" and full of AI cliches, or does it sound like a natural human creator? 3) Platform Optimization: How well does the model leverage LinkedIn-specific mechanics, such as effective hook lines, spacing, relevant hashtags, and viral formatting?
28 experiments14 models tested
GPT-5 MiniGPT-5 NanoClaude Haiku 4.5Claude Opus 4.5GPT-5GPT-5.2Claude Sonnet 4.5Gemini 2.5 ProGemini 3 FlashGemini 2.5 FlashGemini 2.5 Flash-LiteGPT-4.1GPT-5.2 ProGemini 3 Pro
Author’s Rankings
Conclusion
Key Findings & Insights: 1) Consistent Performance Across Scales: I was surprised by how similar the results were across various architectures. No model completely failed the task, showing that basic marketing copywriting has become a baseline capability even for smaller models. 2) Brand "DNA" & Styling: Models from the same provider (e.g., Google vs. OpenAI) showed strong internal consistency, likely due to shared training datasets. This highlights a crucial insight: every provider has a distinct "native style." As a user, you must either adapt your prompting to this style or choose the provider whose default "vibe" best fits your specific task. 3) Gemini’s Versatility (S-Tier): Gemini 2.5/3 Pro stood out by providing multiple options tailored to different audiences, along with unsolicited (but helpful) publication advice. While providing extra options wasn’t explicitly requested, the quality of the "publish-ready" variants made it the winner for me. It transformed from a simple generator into a strategic assistant. 4) OpenAI’s Non-Linear Progress: GPT models performed well, but there was no clear correlation between model "power" (e.g., GPT-5 vs. GPT-5.2) and result quality. Newer or larger didn’t always mean better for this specific LinkedIn use case. 5) The Claude "Emoji Trap": Despite being my go-to for coding, Claude struggled here by overusing emojis. This created an immediate "AI-generated" red flag. While fixable with more instructions, this experiment exposed Claude’s default tendency toward a "robotic" marketing tone compared to the others. Final Verdict: Gemini currently leads for professional social media content that feels natural and offers strategic variety, while other models require more rigorous "human-like" fine-tuning in the prompt.
0 comments