Tools don't use themselves

In my last email I sent out a behind the scenes look at the in-progress Shift Nudge Claude Code training.

I asked what felt unclear, frustrating, over-hyped, etc.

A HUGE amount of responses came in.

There was a common theme across them all.

on the plane ride home from Costa Rica last weekend

I did my best to put them together and share some thoughts…

(quick peek of the trip here on IG)

1. Where do I even start and what is the actual workflow

People want order of operations.

What happens first, next, and what carries the work forward.

Remember that Claude Code and any AI for that matter is a TOOL.

Please make sure to keep this engrained in your mind.

This fact alone will help you cut through the hype, frustration, and confusion.

Tools don’t use themselves.
A hammer doesn’t build a house.
A wrench doesn’t fix a car.

Tools can be used poorly or they can be used with excellence and precision.

Plan Design Build Refine Review

These are the 5 main phases that you can loop through while using Claude Code.

These phases aren’t novel by any means, but it’s important to remember that this is the natural process of designing and building anything.

I genuinely believe the design process now is building an idea quickly, running a prototype, and figuring things out on the canvas, each one informing the other.

The workflow lives in the back and forth.

2. What belongs in Figma, what belongs in Claude and code?

Figma carries exploration, relationships, systems, and visual judgement.

AI carries translation, synthesis, momentum, and implementation help.

Code carries the ultimate truth.

And now code can start carrying that truth earlier than what we are used to.

A static frame, a component page, and a clean Figma file, all still help.

The live build starts answering a different class of questions.

Motion, state, edge cases, and real behaviors show up in the live build.

That is why intent gets so important.

Playing around with the build reveals bugs and edge cases so much faster than any canvas artifact possibly can.

3. How do I get better output through structure, context, and clear prompts?

Structure wins here every time for me.

Strong prompts help, but a strong system helps a lot more.

I’m constantly reaching for building structured context, references, specs, screenshots, project maps, PRDs, and scope instructions.

The more you can solidify the structure of what you know about your project and the documents that can best represent your knowledge, the better structural context you’ll have for working with AI on all the things.

Imagine building a sand castle on the beach.

You’re pulling sand with both hands towards the center so you can build what you want.

That’s how I think of engineering context for AI.

Gathering material, shaping, creating enough mass for the thing to hold.

Then you build.

Here are two real examples from the actual project I built for the Claude Code training.

This is the claude.md that Claude Code will auto-generate for you, like a “rules of the road” type document.

But this is the project_map.md a full "map of the city." In the Claude Code curriculum I walk through in great detail how to create one for your project.

I also ask Claude regularly, “Do you feel 100% confident to do XYZ?” and if not, what are the gaps, what questions can I answer to bring up the confidence level, etc.?

This doesn’t work perfectly every time but it gets you much closer to having a helpful AI tool than without.

4. How do I use AI while preserving taste, judgment, and craft?

By having it do smaller tasks for you.

Use it as a scalpel.

Don’t rely 100% on AI getting it right with a single prompt.

You still need your own brain and the skills you’ve developed more than you need a single “skill” file.

Building, maintaining, and steering context in the form of markdown files becomes a very valuable human skill when working with AI.

But if you don’t have a vision for what you’re building and just prompting and crossing your fingers, I don’t think you’ll have much success.

I’m working really hard to layout context engineering in an approachable way that will work for any AI model, not just Claude Code.

5. How does this hold up when design and development start collapsing into the same process?

These disciplines still exist but they’re blending.

I’ve long thought the best way to design a product is to get it to 75-80% “complete” and finish the rest in code.

There are things you just can’t account for in canvas-only.

I love the canvas and I will still use it as a part of my process.

But code and “feeling the design” in the browser or on the device just tells you so much more than you can hypothesize about in your design tool.

The build should inform the design and now with AI it’s much faster to do this.

The design shapes the build, the build shapes the design, back and forth.

Now the designers who haven’t coded before—using AI—can reach much further into development.

They can shape behavior, structure, state, motion, edge cases, and implementation decisions with a level of proximity that used to belong to a much smaller group of people.

Closing thoughts

I appreciate all of the thoughtful replies to the Claude Code curriculum email last week.

All of these replies are directly impacting the work I’m doing to create incredibly approachable curriculum around wielding the power of Claude Code—or any AI model for that matter—in your professional work.