Does your AI feature use memory nobody asked for, say something is saved when it isn't, or hide where its answers come from?

I review one AI feature and show you where it loses people's trust, then design the fixes: clear consent, visible sources, honest status and approval where it matters.

What you receive • A step-by-step walkthrough of the feature as a person would use it, recording what it says, does and claims at each step. • Prioritized findings, with what I observed kept separate from what I'm assuming. • A trust map: where the feature asks permission, shows its sources, states its status and hands decisions back to the person, and where it skips them. • A redesigned flow with low-fidelity wireframes for the top fixes. • Wording for the honest moments: "I couldn't find that", "not saved yet", what memory was used, and what the AI didn't check. • Acceptance criteria your team, or your AI coding agent, can check the fixes against.

How we work We start with a 30-minute call about who uses the feature and what worries you about it. I use the feature myself, step by step, in a test account or from your screen recording. Then I send the written report and walk you through it on a second call.

What I need from you • Access to the feature: a test account, a staging link, or a screen recording of the main flows. • Who the feature is for, and the one or two things you most want it to get right. • Any complaints, support tickets or feedback you already have, with personal details removed. • One person who can answer my questions during the week.

Scope This is a design review of one AI feature or flow. It is not a code review, security audit, privacy or legal compliance check, or model training or prompt tuning, and I don't build the fixes. I lead the review and the design decisions, using AI-assisted tools where they help.

Options • Quick Read ($450, 3 days): one flow (about 10 screens or a 10-minute recording), prioritized written findings and a 30-minute call. No wireframes. • Full Review ($950, 1 week): everything listed above. • Review + Behavior Spec ($1,800, 2 weeks): the Full Review, plus a behavior spec for the fixes (states, wording, acceptance criteria) and one re-check of the fixed flow once it's built. The listed price is the Full Review. Message me to book one of the other options.

Work I've done in StephOS • Right answers, uninvited memory. In a synthetic test of 24 questions across two versions of StephOS, the assistant recalled 18 of 18 facts correctly, but 2 of 4 unrelated questions still drew on memory. I treated that as a consent problem, not an accuracy win. The assistant now reads the recent conversation and asks when something is unclear. That release shipped October 19, 2026. • One home for memory decisions. Searching on this Mac and sending to a cloud model are separate permissions in StephOS; allowing one never implies the other. Turning memory off, keeping one chat out, and skipping just the next message are three distinct choices, all in one place. (StephOS, October 2026) • The app states the status, not the model. An assistant once told me a saved idea needed approval to last. It didn't. Now the app, not the model, says what's saved and what's approved. (Found in StephOS testing; documented by October 5, 2026)