Claude has completely changed my life. Not in the way that you think. Six months ago, I trained Claude to be a personal trainer and I’m stronger than I’ve ever been.

Before, I relied on group fitness classes. I initially gravitated towards fitness classes because they removed any mental energy in deciding a workout plan and they kept me accountable. But I felt like I had plateaued, and as a burn-conscious early-stage founder, the cost of the group classes was getting expensive.

So, I popped open a “Personal Health and Wellness” project in Claude. I gave Claude a couple of fitness and weightlifting podcasts I listen to and developed skills against those, provided it with my overall fitness goals, and asked Claude to build me monthly workout plans at the gym prescribing specific lifts, reps, and progressive overload over time.

For six months, Claude coached me through every lift and every set, which I would log in real time. At the end of every workout, Claude would build a lift log that summarized the workout that day. Here’s where things got annoying: Claude couldn’t update the lift log. It was a file within the project and shockingly, updating it was something Claude couldn’t do. Clearly we hadn’t reached the singularity. As a workaround, I needed to save the new lift log as an .md file locally on my phone, go into the project, delete the previous lift log, and upload the new one. I did this process four times a week for months. I just wanted Claude to do the thing, but it couldn’t. I had to do it, and it felt like one of the most manual parts of my day.

How does this relate to Gigi?

We see the frustration in our customers every day when Gigi can’t just do the thing, and naturally these create gaps that we need to close to delight our customers. For example, when a customer was happy with Gigi’s output from an ad hoc chat, they would ask Gigi to create an ongoing task so they could recreate that initial prompt and output on a defined cadence. Like my experience with Claude, this wasn’t something Gigi could do. Instead we had to instruct our customers to copy and paste the prompt and recreate it as a task. They were able to do it, but it was annoying, and honestly, some of our customers never followed through on task creation even though it was a couple of clicks away.

Allowing Gigi or Claude to just do the thing takes work. Gigi knew what a task was. She could write one out perfectly in chat. What she lacked was a hand. Our team had to give her three things: a tool she could call to create the task, a guardrail that routes the call through the same creation flow the UI uses so the task is built correctly, and a receipt in chat so the customer can confirm it exists. Same three things for Claude and the lift log. Someone at Anthropic needed to enable file updating within a project as a tool that Claude could call and take action on. They did. The lift log now updates itself.

For both Anthropic and Gigi, it took an intentional choice to do the unglamorous work. This work wouldn’t appear in a capabilities chart, an eval breakdown, or a splashy demo video. But the output of that work is what we all now expect from agents: the agent just does the thing.

Mitigating annoyances and frustrations is the quickest path to widespread adoption. We track these annoyances religiously: every Gigi interaction ends with a thumbs up or thumbs down. Today a customer gave a thumbs down because Gigi couldn’t run the task he created in chat on the spot. By default the task would run tomorrow morning. He just wanted Gigi to do the thing. Six months ago I wanted the same from Claude, and now the lift log updates itself. His task will run on the spot soon enough, and the next gap will be smaller and faster to close than this one.