genouk

Inspiration

Prompting is everywhere. Everyone uses it, and it's hard to be competitive in the modern world without it. Fully understanding and wrapping your head around syntax also takes way too long, especially if you're a beginner. That's why, rather than shunning prompting, we decided to embrace it. The problem we kept seeing is how people fire vague prompts, get garbage back, and end up blaming the AI. The problem isn't the model, but the prompt. So, we made an editor that coaches you to prompt well before you even submit anything, and keep track of everything across sessions so you never start cold.

What it does

Genouk is a VS Code extension available on the store, it acts as a sidekick (with a mascot!) that lives in your editor while you work.

Prompt review & rewrite - paste a prompt and Genouk grades it 0-100 against things that actually matter, then it rewrites it into a production-grade prompt you can paste and run. It's grounded in your repo, so it has context of your entire project through all your files, and the rewrites name real files and frameworks.

Session Planner - give Genouk a goal, and it'll turn it into a structured, editable task that lives in the sidebar or as a popout. The session planner can also easily be linked with Linear, allowing sharing of tasks with other task members.

Repo tours - Genouk generates a live guided tour of the architecture and important features of any fresh repo you may open, allowing you to gain context on where everything is, and get a headstart on your work.

Cross-chat memory - Genouk writes a digest of each session and syncs it into a managed block so the next AI chat loads what the last one did.

How we built it

We used typescript and bundled to have an extension host, a webview, and a standalone MCP Server.

The sidebar and pop-out both render the same React bundle and stay in sync through a shared function, so edits in one window push to the other.

The extension runs on Vultr Serverless Interference (DeepSeek-V4-Flash) with the API key stored within secretstorage.

Feature modules are all separate functions but all share one function for repo grounding.

Challenges we ran into

Early prompts were all giving scores of 70 and were useless, we rebuilt the rubric to start from 0 and only award points for real signals.

The recall for the cross-chat-memory wasn't working and depended on the agent calling a tool which was flaky, so we switched to block editors that auto-load every session.

Our sessions were being cut off due to low usage limits from Gemini and Groq, so we switched to Vultr, which didn't just make the responses better, but also improved speed and usage greatly.

Accomplishments that we're proud of

We built a blunt reviewer, not just a yes-sayer that gives no honest feedback

We created cross-chat memory that carries context between AI chats without tool calls.

What we learned

The problem with vibe-coding is rarely ever the AI model you use, it's rather the prompt quality and carried context you have. We also found how important honest and strict feedback is to creating greatness.