Yesterday, I told yal about my use of Claude to build a React web experience that will be embedded in a mobile app. It worked, but the UI wasn’t quite where I wanted it to be.
Today, I tried a different workflow.
I used Google Stitch to create the interface designs and prototype, prompting it with our brand colors and visual guidelines. Stitch produced a much stronger UI that was closer to what I had envisioned.
Then I connected Stitch to Claude through an MCP server. I asked Claude to pull the designs directly from my Stitch project and turn them into a functional React experience.
The workflow looked like this:
• Design and prototype the UI in Stitch
• Use MCP to make the project accessible to Claude
• Have Claude retrieve the designs and write the working code
This division of labor makes a lot of sense: use a specialized design tool to shape the experience, then use Claude to handle the engineering work required to make it functional.
Instead of asking one AI tool to do everything, connect the right tools and let each one do what it does best.
Today I tested Claude’s new Fable 5 model to see what the hype was about.
I asked it to build an interactive React web experience designed to be embedded in a mobile app. The page needed to help users explore local establishments and the deals they offer.
The source data already existed in a Google Sheet, including Google Drive links to each establishment’s images. Claude was able to:
• Pull the relevant data from each column
• Download and compress the linked images
• Associate each image with the correct establishment
• Build the interactive React page
• Apply our organization’s brand guidelines automatically using an existing skill
That last part stood out. Because the brand guidance was already packaged as a reusable skill, I didn’t have to repeat those instructions in my prompt.
This is the kind of AI workflow I’m excited about: turning structured data and existing assets into a usable product experience with far less manual work.
I’ll keep testing Fable 5, but so far, so good.
hen you work in an organization with multiple calendars for different teams and purposes, finding the right information can be frustrating. Turning every calendar on creates a jumbled collage of colors and overlapping events. Searching manually often means clicking through multiple screens and guessing the right keywords.
With the connector, I can simply ask Claude:
• What office events should I pay attention to this week?
• When is the office potluck?
• When is payday?
• What events are happening this month?
It can also help retrieve information from past events.
For example, I recently needed to contact a project manager from a vendor meeting, but I couldn’t remember his name or email address. I asked Claude to find my most recent meeting with the company and list the participants and their email addresses.
Boom. I had the information in seconds, with no need to dig through calendars, emails, or multiple interfaces.
The bigger opportunity is bringing this capability into an organization’s internal chatbot. By securely exposing organization-wide calendars through MCP, employees could use natural language to quickly find company holidays, office events, kickoffs, important meetings, and other time-sensitive information.
Instead of navigating calendars, you just ask a question and get the answer.
We might be onto something.
Every Monday morning, I prepare a departmental and active-project update for the IT leadership team, which I’m also part of.
The goal is to give all of us a clear, week-over-week view of progress, priorities, and anything that needs attention. Traditionally, creating this report takes a while because the information is scattered across Asana, Jira, Slack, email, and other sources.
Today, I ran a scheduled task in Claude Cowork that I had set up to handle the first draft automatically.
It reviewed the different sources, gathered the previous week’s updates, and formatted everything into the high-level bullet points I prefer. I still needed to correct a few details and make some edits, but I’d estimate that about 80% of the report was completed for me.
For a first run, that was a meaningful time-saver.
I also gave the task feedback, so next Monday’s version should be even better. That’s the part I find most useful: the workflow doesn’t have to be perfect on day one. I can keep refining it until the recurring task becomes faster, more accurate, and increasingly hands-off.
First generation complete. Now we keep improving it.
We continue to keep cooking. 🔥
Yesterday, I used AI for one of those low-glamour tasks that quietly eats up time: processing stakeholder feedback on a feature spec.
The spec was in a Google Doc that had been shared for review. Instead of opening the document, scrolling through every comment, and cross-referencing each one with the right section, I asked Claude Cowork to open it through its Chrome extension and organize the feedback for me.
It returned a structured summary showing:
• Who commented
• Where the comment appeared
• What they said
• Any replies, including mine
The useful part wasn’t just capturing the comments. Claude connected each one to the section of the spec it belonged to, so I immediately understood what was being challenged and why. It even flagged a truncated comment rather than silently skipping it.
Once everything was organized, the patterns became obvious. Multiple stakeholders were pointing to the same gaps: collecting a phone number, creating a support ticket alongside the real-time alert, and clarifying the SLA wording.
I then used AI to incorporate those changes back into the spec instead of updating it manually, line by line.
The biggest productivity win wasn’t speed. It was reducing the cognitive load of constantly switching between the spec, stakeholder comments, their intent, and the revisions.
Now we’re getting somewhere!
Yesterday I used AI as a QA partner for a new department chatbot we’re testing.
Using Claude Cowork with the Claude Chrome extension, I had it navigate the chatbot like a real user: clicking through pages, asking test queries, checking flows, and documenting what worked.
Then it generated a report with:
• what performed well
• what needs improvement
• bugs it found
• steps to replicate each issue
• suggested ways to fix the bugs
The best part: I tested the steps myself, confirmed I could reproduce the bugs, then asked Claude to create the Jira tickets. Because it already had project context, it knew the right teammates to assign them to and included the suggested fix or recommended path to resolve each bug in the issue details.
A few clicks turned into test coverage, a findings report, validated repro steps, recommended fixes, and Jira bug tickets.
Now we’re cooking!
Today I used Claude Cowork with my Jira plug-in to turn sprint review action items into actual work items.
I gave Cowork the notes, asked it to create the right Jira items, and had it bring everything back to me for review first. Once I approved, it categorized the work correctly as user stories, spikes, feature requests, or bugs, assigned each item to the right person, and posted it into the right project.
That used to be a manual task: rewrite the notes, choose the issue type, create each ticket, assign owners, and make sure everything landed in the right place.
Now I can review the output, make a few clicks, and move on to higher-value product work.
This is the kind of AI workflow that actually matters. Not replacing judgment. Removing the busywork around it.
Today was a perfect example of how practical AI saves real time.
A vendor asked me for contact info for someone I worked with years ago. I vaguely remembered the person’s name, but not their email address or current details.
In the past, that would have meant digging through Gmail, trying different search terms, opening old threads, and piecing the context together manually.
This time, I went straight to Claude, which I already had connected to my inbox. I gave it the rough context: who I was looking for, what the conversation was probably about, and why I needed the info.
Within seconds, it found the person’s name, title, and email address. I copied it, sent it to the vendor, and moved on.
That is the kind of AI use case I care about most: not flashy, not theoretical, just immediate workflow acceleration.
Speed is the name of the game.
Seconds later: name, title, email.
Copied. Sent. Done.
Speed is the name of the game.
Today I used Claude Cowork to set up an MCP server that connects to Google Analytics for one of my products, an AI chatbot. The setup did not take long, and now the product metrics are available through natural language instead of fixed reports.
That changes the workflow. I can ask questions directly, explore usage patterns, and generate ad hoc graphs or tables without waiting on a manual analytics pull. It is a small experiment, but it points to something bigger: analytics becomes more useful when it moves closer to the questions we are already asking.
AI Compute as a Service (AICaaS) has the potential to become a major growth category, particularly for companies looking to scale AI inference efficiently and cost effectively. As demand for high performance compute continues to accelerate, orbital data centers could represent the next frontier, offering expanded capacity, lower latency across global networks, and entirely new infrastructure possibilities for AI at scale.
AI Compute as a Service (AICaaS) has the potential to become a major growth category, particularly for companies looking to scale AI inference efficiently and cost effectively. As demand for high performance compute continues to accelerate, orbital data centers could represent the next frontier, offering expanded capacity, lower latency across global networks, and entirely new infrastructure possibilities for AI at scale.
As the recently expanded partnership with @AnthropicAI demonstrates, @SpaceX is offering AI compute as a service at significant scale.
We are in discussions with other companies to do the same.
Over time, especially with orbital data centers, we expect to serve AI at extremely high scale.