The role of comprehension in AI workflows
Issue 230: Understanding before progress, and shared knowledge between humans and sentient tools
Whether we like it or not, Artificial Intelligence will continue being the dominant headline over the next few years. Time can only tell if this is a new paradigm shift or something that fades way. For me, it’s the moment to shape it. For software makers, the factory reset encourages everyone to be a beginner again.
Headlines tell different stories. Every week there is a new breakthrough that changes how we navigate the displacement variance. This past week, the Chinese AI Startup launched DeepSeek, a model that arguable rivals Open AI’s while at the same time Open AI ships a preview of their agent—Operator. On the flip side, the adoption of AI has not made progress as people though. We were told AI would replace us an instead we received sparkle icons and lots of AI slop.
If you’re familiar with Crossing the Chasm, this is not a surprise. AI tools and workflows are in the Early Market phase. There are a small group of champions in the workplace geeking out about it, like the first person who got access to Microsoft Excel while others were doing more manual number crunching.
Behavioral defaults are difficult to change, especially in the work place. This is why comprehension, the ability to understand something, is a prerequisite to adoption. Despite the new generative features and automation of AI, the ability to understand the constructs of what it’s attempting to do in order to achieve good results. For writers, understanding linguistics will allow creating more believable languages in their fantasy novel. For designers, having a strong foundation in drawing fundamentals create better understanding of building interfaces. Understanding the constructs of the current technology might give you better results. It’s also important to note comprehension is critical for the human being and the AI; a sentient tool you collaborate with.
First, let’s cover human comprehension. When I was a design instructor, we were taught to check for understanding. The check point ensured students could ask questions if a concept wasn’t clear. If you don’t check for understanding, the likelihood of a student becoming overwhelmed due to lack of comprehension increases. AI workflows allow opportunities to provide insights in the UI that the end user can learn—increasing the chance of comprehension. The best authoring tools improve comprehension as users engage with them. For Webflow, it was “No-code, know code”; for Replit, the Agent’s UI signals foster learning by showing users its processes.
In Replit’s Agent, the Progress pane tracks application development. While initially unfamiliar to novice coders, the agent’s explanations help users gradually absorb knowledge, building understanding and proficiency in software development.
Next, let’s discuss AI comprehension. My personal focus working in AI is less about building AGI. Instead, I view AI as sentient tools and teammates—something unique instead of an exact mind like a human being. These sentient tools will only be as effective as the human being prompting and giving direction. For example, saying, "Protect humanity at all costs" to a neural network AI like Skynet results in it protecting humanity from the biggest threat to humanity...humans. Whoops.
The current reality of AI tools is it requires clear constraints and context. Without those, the AI might do something you did not inspect. I’ll give you an example. When I use a natural language interface such as Cursor or Claude Artifacts, there are several instances when I instruct the AI to make a simple task such as adding a new page on my website. At times, the AI might re-write or remove other elements of my project I did not intend it to.
This will improve over time. For now, checking for understanding with the AI will save you time in the long run.
When you combine the feedback loop for AI and human comprehension, it causes more instances of what I call the AI game of context telephone. Recently, a person on Twitter posed a screenshot of a real Apple Intelligence summary of his (ex)girlfriend breaking up with him.
AI summaries synthesize the text without context. Summarization and automation without comprehension can result excessive AI content. This can happen with meeting transcripts that don’t have the context. It also happens when Person A uses Generative AI to write a long professional email and Person B summarizes it in one sentence, resulting in context and understanding that’s missing.
The opportunity I see is finding affordances in AI experiences to check for understanding.
The desired result is an harmonious loop of human and sentient tools learning how to work together through feedback loops from prompt to execution.
Hyperlinks
Man learns he’s being dumped via “dystopian” AI summary of texts
You can just build things → It’s so cool to see designers shipping software again
Congrats Horse Browser on two Golden Kitty nominations: Pascal as maker of the year and Bootstrapped & Small Teams → Give them a vote
Job opportunities
First Product Designer at Visual Electric
Lead Content Designer for Confluence AI at Atlassian
Lead Designer, AI at Atlassian
I am an investor
One nice benefit of reasoning models that log out their chain of thought is that you can read exactly how the AI understands your request and context. It can really clarify how well you've communicated your intention. (If only communicating with other humans were so easy...)