Project chapter
AI-powered writing assistant
Managers avoid giving feedback because they don't know how. I took a hackathon prototype to a production LLM feature that was in-app, secure, and built around expert prompts, cutting feedback cycle time by 10% and driving 65% feature adoption.
Managers who can't give feedback don't give feedback
Managers at small and growing companies often get promoted from IC roles without training. When it comes to feedback cycles, many feel blocked. They can identify problems but can't articulate them clearly, fairly, and actionably. The result: delayed or low-quality feedback, damaged development cycles, and frustrated reports.
At the same time, many managers were already using ChatGPT to write feedback outside our product. We had a window to deliver something better: secure, context-aware, and built into the flow.
Small companies don't have HR training infrastructure. Promoted ICs get stuck when asked to give structured, actionable feedback.
Our differentiation couldn't just be “AI feedback help”. We had to be more valuable than the free alternative already in their browser.
B2B HR customers are compliance-conscious. Any AI feature had to be secure, with no employee data used for model training.
From hackathon prototype to production feature
I started with a hackathon prototype (“Feedback Buddy”) that incorporated multiple feedback methodologies: Radical Candor, NVC, STAR. It was powerful but complex. Customer interviews pointed us toward something simpler: in-the-moment, in-app assistance that felt invisible and immediate.
- 01Research & discovery
Interviews with senior HR admins and managers to map feedback friction. We found managers already using AI, but outside the product. That validated the opportunity and helped us define what “better” looked like.
- 02Prototyping & prompt engineering
Hundreds of prompts tested to move from complex multi-framework output to something lightweight and usable for any manager. The key insight: quality over sophistication. Simpler prompts, better for most audiences.
- 03In-app implementation
Built directly into the rich text editor with no switching tabs and no context loss. No employee data used to train the model. Public AI documentation prepared to answer compliance questions proactively.
- 04Beta measurement
Launched under beta label with in-app feedback mechanism. Tracked regeneration rate, applied rate, and qualitative input. Used real feedback to refine prompts and UI iteratively.
In the moment, in the editor
The assistant appeared directly in the feedback writing flow, not as a separate tool or tab. Here's what a manager would experience:
Alex delivered the Q3 report late and the quality wasn't what we expected. They need to improve on deadlines.
In Q3, Alex's report was submitted after the agreed deadline, and the depth of analysis didn't meet the standard we'd discussed. I'd like to work together on a plan that helps you build stronger time management habits and a clearer framework for structuring analytical work. I'm confident we can get there.
Switching to ChatGPT meant losing context. The manager had to copy-paste, re-explain the situation, and paste back. Our solution was zero friction: the assistance happened where the feedback was being written.
HR teams couldn't adopt an AI feature that used sensitive employee data for model training. By explicitly not doing this, and publishing our AI policy, we removed the biggest objection upfront.
Usage that spoke for itself
What I know now about shipping AI features
Our most technically impressive prototype (multi-framework) was less useful than a simpler, single-mode assistant. Match the cognitive load of the feature to the moment of use.
The biggest differentiator wasn't the AI quality. It was that the feature existed in the exact moment of need. Integration beats capability when the use case is high-friction.
In B2B HR, proactively answering “is this secure?” unlocked adoption from customers who would otherwise have blocked rollout. Don't wait for the objection.