Context Goblin
Blog
-
· Clemens
How to build a startup with AI
We let LLMs research and build autonomously, but refuse to let AI write text for a human reader.
-
· Niklas
Seventeen comments that changed other people's code
We had no benchmark to show, so we pointed Context Goblin at pull requests in repositories we don't own and let the maintainers judge. Maintainers at n8n, Astral's uv, OpenHands and five other projects took our comments into their code, three of our own patches merged, and one maintainer told us his single pass with Claude Code and Codex had missed what we found.
-
· Clemens
Ads don't help you get qualified customer feedback
We spent €200 on Google Ads and Reddit Ads for Context Goblin and got 3,900 conversions and ~80 signups - but the 50 people we reached through friends gave us better feedback than all of it.
-
· Clemens
What we learned launching on DevHunt: 3,351 impressions, 22 visitors, zero signups
We paid $49 to launch Context Goblin on DevHunt and finished the week in second place. The site's front page is dominated by six-figure impression counts from past winners, while the week we actually bought into ran two orders of magnitude below that. Here's the gap, the 22 visitors it sent us, and the check we should have run first.
-
· Clemens
The model is not your security boundary
Prompt injection can't be closed at the prompt layer. Assume it worked, then ask what the agent was holding. What our own audit found, and why the OpenAI and Hugging Face incident is worth reading if you point agents at input you don't control.
-
· Clemens
Everyone reads the whole repo now. Here's what still sets these tools apart
A plain comparison of Context Goblin against CodeRabbit, Greptile, Cursor BugBot, GitHub Copilot code review and Anthropic's own Claude Code Review, on architecture and capability, not on a benchmark we don't have.
-
· Clemens
Your AI reviewer has never read the rest of your codebase
Most AI code reviewers see a patch and nothing else. Context Goblin works out what a change actually touches, in your repo and across the ones around it, before a single agent starts reading.