AI
The Real Costs of Unreviewed AI Code
Vibe coding means letting your AI agent generate all the code and merging without reading it. Scarlett called it
“incredibly and dangerously irresponsible.”
One Twitter user handed Claude a prompt and watched Claude delete everything they had built. Unreviewed AI code also adds insecure patterns to enterprise codebases and creates persistent bugs nobody understands. Simon Willison’s term “vibe engineering” captures the fix: work with LLMs while staying accountable for what ships. Patrik Eriksson’s version is “chiselling.” The label does not matter. The principle is that you own the code.
Choosing Your LLM and Feeding It the Right Context
Not every LLM makes a capable copilot. Scarlett picks on tool-calling capability first: an LLM that cannot invoke tools turns your agent into a glorified chatbot. Context window size and code reasoning come next. She uses Claude Sonnet 4.5, Claude 4, and GPT-4, sometimes routing one as reasoner and another as executor inside Goose.
Context files cut hallucinations. Scarlett’s top entry in her goose.hints file: commit after every change. That builds a granular Git history, so when an AI edit goes wrong you can pinpoint the exact commit and roll back.
MCP: Giving Agents Access to Real Tools
Model Context Protocol is an open standard that connects any supporting AI agent to external tools via natural language. Scarlett’s example: instead of writing SQL, tell your agent “select all the people who attended my talk.” The GitHub MCP Server handles rebases, merge conflicts, and project management without remembering git syntax.
She showed this live in the GitHub MCP Server live demo (08:53): one prompt created a GitHub issue assigned to her account, a second rendered a full interactive project calendar, and clicking “analyze workload” returned per-person task loads from live GitHub Projects data.
Staying in Control with ACP and Chiselling
Agent Client Protocol lets you bring any supporting agent into any editor, so you stop context-switching between a chat window and your code. Scarlett runs Goose inside Zed this way. Staying in the editor is where chiselling becomes practical: you see every change as it lands and can still ask whether it is secure and scalable.
“Working code isn’t finished code.”
Patrik Eriksson’s full version: the kernel of what you wanted is buried under a mountain of AI-generated slop, duplicated functionality, thousand-line methods, useless unit tests. Chisel the excess away. That is the job.
Notable Quotes
it is also incredibly and dangerously irresponsible. Rizel Scarlett · ▶ 2:15
Working code isn’t finished code. Rizel Scarlett · ▶ 12:21
Some models are really better at reasoning through complex code than others, Rizel Scarlett · ▶ 4:49
Key Takeaways
- Commit after every AI change so you can pinpoint and roll back problems without hunting through diffs.
- MCP connects your AI agent to any tool — database, GitHub, Blender — using plain English instead of syntax.
- ACP keeps you inside your code editor while the agent works, making code review a natural part of the flow.