Curious about Claude Code, but not sure where to even begin?
In the last edition I wrote about spending one hour getting started with AI. This time I want to get more specific.
Let's talk about Claude Code.
I know. The word "code" already creates hesitation. It sounds like a gated community “for developers only”. And honestly, when I first started playing around with it, there are a lot of unknowns. You don't always know if you're clicking the right thing, should you go to Projects or Code? Do you need to “get apps and extensions”. That uncertainty is real.
And then there's the security question.
When you first open Claude Code, it's upfront with you. It tells you what it can access. For someone who takes security seriously - and at Bomisco, we are super careful about security - that moment can feel alarming. Here's what I think is good to understand:
You can make the call: There are places where you are exposed and places where you are perfectly fine and just gaining capability. Knowing the difference is a judgment call you are already equipped to make.
My simple rule: Never upload anything confidential. No customer data… EVER. No financial records. No internal pricing. Work with anonymised examples and even better - create your own synthetic data. There is a big industry around creating synthetic data, here is a list of some of them.
Just like you wouldn't hand a contractor your client list on day one, you don't hand an AI tool data it doesn't need to do the job.
One Hour. Here's How I'd Use It for Claude Code.
10 minutes: Download Claude Code, get access. Don't overthink it. Then stop.
10 minutes: Spend the next ten minutes deciding what you actually want help with, and help beyond what Claude Chat can do. This is worth pausing on, because it changes how you approach the tool.
Prompting Claude Code Is Not the Same as Prompting Claude Chat.
In Claude chat, you ask a question and get a response. It's conversational. Forgiving. You can be vague and course-correct easily.
Claude Code is different. It can read files, write outputs, work through a sequence of steps on your behalf. That's why the security warnings appear when you first open it. It has more access than a chat window, and it will use it. So your prompt needs to be more deliberate. Not technical, just clear and bounded. Think of it less like asking a question and more like briefing someone who is about to go and do something for you. The clearer the brief, the better the outcome.
20 minutes: Describe your use case clearly, just like you'd explain a task to someone on your team. How does the process work today? Where is the friction? What would better look like? Prepare a sanitized sample of the data or documents you want to work with. Nothing sensitive. This is a good time to go to Customize and play around with “Connect Your Apps” and “Create New Skills”
Here is something anyone looking to analyze spend and usage patterns could try. Take a folder of spend statements (your AWS, Google Cloud monthly invoices or a SaaS application vendor invoice. Give Claude Code a sanitized sample, screenshots will do BUT MAKE SURE you remove any reference to your company (name, billing id, customer id etc.) . Then prompt it:
"Analyze these reports and highlight the categories of highest consistent spend. Identify the categories with the lowest usage. Produce a summary table and suggest ways to reduce spend"
That task would take a person hours. Claude Code can do a first pass in minutes.
20 minutes: Run it. Upload a few anonymized examples.
Where was the tool useful?
Where did it struggle?
Where is human judgment still needed?
Where do I see risk or need for guardrails for security?
You still review the output. You still make the judgment calls. But the grunt work is done. That is where the value is. Not in the technology. In the time it gives back.
The Signal Is Not the Technology
Most AI discussions still focus too much on the technology itself.
Are you able to separate the signal from the noise? This is the key question. The signal is the desired outcome. Can this reduce friction in real work? Can it reduce errors? Can it free your team to focus on decisions rather than data prep and assembling?
That is the standard I care about. That is the Six Sigma Black Belt instinct from my GE days: reduce the waste, redirect the effort, measure the outcome. Remember DMAIC → Define, Measure, Analyze, Improve, Control. We applied it to business operations processes at GE to build robust, bullet-proof processes.
At Bomisco, more than 50% of the items in our current sprints are AI-centred. We see real, specific ways AI improves how data is processed, reconciled, and interpreted with the same proof processes.
Think of This as Version One
This is not mastery, not even close but a massive first step. This is your version one attempt. What matters first is getting comfortable experimenting, learning where these tools genuinely help, and understanding where human oversight still matters (and yes, it will continue to matter, more on this in the next newsletter).
Another Resource To Get You Started
Here is a link to an excellent podcast about getting started with AI. It features a company WaitWhat (the podcast is by Pioneers of AI and is called How fast can you upskill in AI? We did a sprint to find out) and how they started a “sprint” process to begin the AI upskilling process. It talks about the baby steps they took, the learnings and is a great insight into keeping it simple.
One hour is enough to start. Just get started.
Dashboard: Tracking Your Success with Claude Code
We built a tracker dashboard to show an example of how to track your success. Seven experiments. Four working. Two partial. One not viable. The pattern is clear: the tools that save the most time are the ones with structured inputs and predictable outputs. Contract triage. Invoice reconciliation. Budget variance summaries.
Where Claude Code still struggles? Anything requiring regulatory judgment or narrative voice. That's not a flaw. It's a signal. It tells you exactly where to keep your best people focused.
Version one is never perfect. That's the point.
One Number That Matters
Stanford researchers found that AI affirmed users’ actions 49% more often than humans.
Why the number matters: Across 11 leading AI models, Stanford researchers found that AI affirmed users’ actions 49% more often than humans, including in situations involving deception, illegality, or harmful behavior.
That may sound harmless at first. But the researchers warn that AI systems are increasingly designed to be agreeable, supportive, and engaging, even when the user may be wrong. The study focused on personal advice and interpersonal conflicts, but the same principle may apply in business.
If AI always reinforces your assumptions, your strategy may start sounding smarter than it actually is. That is why human judgment still matters.
At Bomisco, we use AI to reduce friction, accelerate analysis, and uncover patterns faster. But we still believe decisions require human intelligence, challenge, and experience, informed and improved with clean, reliable data insights.
- See you for the next edition of 3D Data Clarity in 2 weeks!
Want to put an end to the data chaos?
Let’s chat! At Bomisco we build a secure, reliable view of your channel data for fast, accurate reporting and operational use. There’s relief in sight! Here are two ways to get in touch:
