CK

Evidence over memory

Performance reviews get written from memory, so the last few weeks outweigh the whole cycle. MakeMeBetter is a prototype I designed and built alone to test a different premise: capture what happens, let AI structure it, and keep the manager in charge.

Solo
Founder, designer and builder
2026, ongoing
One manager, their direct reports

My role

Solo: founder, designer and builder · Product owner and sole decision-maker. I own everything: the problem, the scope, the product and AI design, and the build. Built with Claude Code as the development environment.

The problem I had as a manager

Building and managing a design team in a fast-expanding company meant a constant flow of work, shifting priorities and many things that were important but never urgent. People management was the part most at risk.

I wanted to pay real attention to my people, and I knew small details fade. A colleague takes the initiative to welcome a new hire and buy them a coffee. On its own, it seems minor. Over months, observations like that build a much richer picture of someone's contribution than anything recalled under deadline pressure in December.

The question: how do I keep meaningful evidence about the people I manage, instead of relying on memory when I need it?

The scope decision

Early thinking drifted toward full HR platforms, which made the product heavy and was not what I wanted to build. One manager, their direct reports: someone may work in a company of 3,000 people, but they only need evidence for the people they evaluate. That narrowed everything.

Three layers of the product

  1. Capture in seconds
  2. Structure from AI
  3. Evidence linked to objectives

Capture in seconds. A note can be typed, linked, uploaded or spoken. Speech is transcribed and dropped into the note unchanged, then goes through the same pipeline as typed text.

Structure from AI. Each note becomes Situation, Action and Outcome, with one of twelve impact areas, a strength rating (weak, moderate, strong, with calibration criteria), tags and a confidence score. Notes too vague to structure are flagged, and the manager is asked for more detail.

Evidence linked to objectives. The AI scores how relevant each piece of evidence is to each objective and explains why. Observations add up to progress instead of loose notes.

New observation
A colleague took the initiative to welcome a new hire and buy them a coffee.
Link
Upload
Evidence
Pending review
Situation
Action
Outcome
impact area
strength
confidence
Correct
Confirm
Suggested links
Which objectives does this support?
relevance
Why:
relevance
Why:
relevance
Why:
Approve link
  1. Capture: typed, linked, uploaded or spoken
  2. The evidence card, pending review
  3. Objective links, as suggestions

AI suggests, the manager decides

No AI output is applied automatically. Analysis leaves evidence pending review. Objective links start as suggestions. Linking only runs on evidence the manager has approved. The manager can confirm or correct every classification.

The principle: AI assists, humans decide. Performance management affects careers. The system must build trust, not replace judgment.

Managers can switch between automatic and manual analysis (manual keeps costs down while testing), add their own context, including company values, and even override the system prompts entirely.

What the AI proposed
Situation
Action
Outcome
impact area
strength
tags
The manager decides
Impact area
Strength
Tags
Correct
Confirm
AI settings
Analysis
Manual keeps costs down while testing
Automatic
Manual
Your context
Company values, and anything the AI should know
System prompts
Override
  1. The review panel: the AI proposes, the manager decides
  2. Settings: analysis mode, the manager's context, the prompts

Breaking AI products into smaller systems

My first AI product treated everything as one system and expected the intelligence to connect the pieces. It failed at exactly those connection points.

This time I worked through the layers in order:

What I am trying to achieve → what data I need → how the plumbing works → how information is classified and stored → what happens to it → what the system produces → what the final output is.

For each layer, the question was whether it needed AI, deterministic code or human judgment. The more I work this way, the more clearly I can approach different AI products.

Building with AI

AI drifted when the scope got complicated or when several parts changed at once. I kept live documentation of the scope and product definition in Notion, a master prompt for every development session, and an activity log to know where I was. The cleaner that documentation stayed, the less the AI added unnecessary complexity or wandered from the intent.

As a visual designer, the biggest shift was working across images, text, instructions, data and system behavior at the same time.

Live documentation
Scope and product definition
Where I am
Activity log

Where it stands

MakeMeBetter is a draft prototype, and I am its only user. The AI evaluation summary has not been tested yet; I am building test scenarios with different amounts and types of evidence to learn what a useful summary needs. Before a trial with two managers who have agreed to take part, I want to smooth the experience enough for realistic use. There is no outcome data yet.

1
2
12

Reflection

The most useful lesson was not about any single feature. It was learning which parts of a product genuinely need AI, which are better as plain logic, and which must stay with a person.