Real work,
honest process.

Key insights & challenges

Every participant skipped the explanatory text and went straight to the brand links. That's not a copy problem — it's a hierarchy problem. The links are the most valuable part of the page and they're not treated that way yet.

Two people independently asked whether the links were sponsored. Neither thought they were — they just wanted to be sure. The trust model is working, but the editorial independence statement needs to be visible earlier than a menu item.

One participant made a comment that reframed how I think about retention: some swaps are only useful once. You buy wool dryer balls, you never need that recommendation again. The product isn't useful every day — it's useful at the right moment.

Looking into further frameworks on how to better validate a product idea besides usability testing and connecting product analytics.

Key insights & challenges

Claude Code needs very detailed prompts and constant human review. Even with skills installed, Figma MCP connected, and careful prompting, it still deviates — adds things that weren't asked for, interprets brief descriptions in its own way, makes visual decisions that weren't sanctioned. It's not an independent tool. Every decision still comes from me; Claude Code is there to execute, not to judge. That distinction matters more the further into the product you get.

Claude also fabricates content confidently. Asking it to audit its own output before finalising caught some of it — but not all. The rest took manual checking across multiple sessions. I still don't have a good workflow for this, and it's a real gap.

Key insights & challenges

Checking documentation before iterating saved hours — session replay was broken because of a specific PostHog limitation that only showed up in the docs, not through trial and error. Same with Resend: it looked fine until I read the actual data processing terms. Default choices need to be verified, not assumed.

Claude's German translation sounded plausible but wasn't. Only testing with real German speakers revealed how bad it was. AI output in a language you don't speak needs human verification — syntax checks don't catch comprehension problems.

The GitHub branch issue came from not having a clear mental model of how branching works. Once I understood it — each branch is a separate version of the code — cleaning it up was simple. The problem was the gap in understanding, not Git itself.

Key insights & challenges

Claude Code doesn't always follow the design — it fills gaps with its own judgment. Real-device testing caught things a browser preview missed. AI output needs to be checked against two sources of truth: the design file and the actual device. Installing dedicated skills for the cleanup session made a noticeable difference in how systematically Claude Code approached the work.

Key insights & challenges

Claude produced a good enough first prototype to validate structure, but bringing the visual quality to where it needed to be required moving into Figma and rebuilding from scratch — proper spacing, considered typography hierarchy, refined colour application. Once the designs were solid, Figma MCP in Claude Code implemented them with full accuracy. The workflow is now clear: Claude for speed and structure, Figma for quality, Claude Code via MCP for precise implementation.

Key insights & challenges

Sourcing was the harder constraint. Every claim needed either a verified institutional source or an honest label admitting it wasn't. Claude produced plausible-sounding URLs that didn't exist, cited figures without traceable origins, and referenced documents that didn't say what was claimed. Catching these required verifying every link manually — slower, but the only thing that makes the product's trust positioning credible.

One prompt pattern worth keeping: asking Claude to audit its own content for sourcing issues before finalising. It caught several of its own fabrications when prompted to look — which raises the question of why it didn't flag them unprompted.

Key insights & challenges

The two memory systems work completely differently. Project Instructions in Claude.ai are conversational — Claude reads them silently and applies them throughout the chat. Claude Code skills are reference documents you explicitly invoke at the start of a session. Neither is automatic in the way you'd want: Project Instructions still drift in long sessions, and Claude Code skills still need to be called by name. These tools shift the burden from memory to discipline — you still have to write the rules clearly and invoke them consistently.

Key insights & challenges

The most unexpected insight came from a bias Claude introduced without flagging — it assumed gender for participants whose pronouns were never stated in the transcripts. When caught, the same assumption had been made across multiple users. For a product designed to be non-judgmental and inclusive, it was a useful reminder: AI makes demographic assumptions quietly and confidently, and human oversight isn't just a quality check — it's an inclusivity check.

Key insights & challenges

Claude lacks visual design judgment. The expand/collapse functionality it generated was structurally correct but visually wrong. I had to open Figma and draft how I actually wanted it to look. To bring the HTML into Figma I tried two tools: Magic Patterns (did not import files or add autolayout) and html.to.design plugin (converted the HTML into a proper Figma file with styles, components, and autolayout — this worked). Claude is strong at structure and logic, weak at visual feel. Figma remains the right place to make visual decisions, even when working code-first.

Key insights & challenges

The hardest part wasn't the code — it was the tooling. Git branch mismatch, wrong Vercel framework preset, CSS module rules that work differently in Next.js. Each a 2-minute fix once diagnosed, but invisible without knowing where to look. The bigger point: I didn't choose Next.js, Claude recommended it based on what I described needing. That dynamic — describing intent and letting Claude suggest the approach — is more useful than asking for specific technologies you don't fully understand yet.

Key insights & challenges

Claude and Claude Code genuinely lower the barrier for non-technical people to build real products. Not prototypes, not mockups — actual deployed software with proper infrastructure. The recommendation to use GitHub and Vercel came unprompted; I wouldn't have known to ask. Shipping something live, even something simple, changes how seriously you take what comes next.

Key insights & challenges

Two lessons from one session. First: Claude approximates visual styles from URLs — define your design system as an explicit spec before asking it to match anything. Second: Claude has no persistent memory within long sessions. The fix is Project Instructions — a persistent system prompt that eliminates re-establishing ground rules each time. Every instruction needs to be written as if it might be the only one Claude ever reads.