You know that sinking feeling when you’ve spent ten minutes typing a perfectly crafted prompt, only to watch the AI generate a plan that’s mostly right but has one critical, bone-headed mistake? You try to fix it with another prompt, and another, and suddenly you’re trapped in a game of telephone with a machine that just won’t get it.
I’ve been there. So has the developer behind open-plan-annotator, a local annotation UI that lets you literally cross out bad ideas, add comments, and delete steps directly in the browser. No more back-and-forth chat. Just a strikethrough.
Conversational UIs are a trap for complex AI workflows. The future of human-AI interaction isn’t better chat—it’s a regression to spatial, direct-manipulation interfaces designed to edit AI cognition.
Here’s the uncomfortable truth: as AI agents become more autonomous, the bottleneck shifts from prompt generation to plan verification. You don’t need to write a better prompt. You need to see the AI’s reasoning, grab a red pen, and tell it exactly where it went wrong. That’s not a conversation. That’s editing.
Think about how you work with a human colleague. You don’t email them a paragraph and hope they parse your nuance. You walk over, point at the whiteboard, and say, “Scrap that line, start here instead.” Why should AI be any different?
“I’ve been using this tool every day for months,” the creator says. “It’s become one of the few side projects I legitimately enjoy using.” He’s not alone. The GitHub repo fills with users who find simple annotation—strikethrough, comment, delete—more ergonomic than any native agent harness.
One skeptic asks: “Do annotations persist after plan approval? Can they influence future plans?” Good questions. But they miss the point. The real power isn’t memory—it’s the act of editing. When you cross out a step, you’re not just correcting a plan. You’re training the AI in real-time, without the friction of natural language.
This is the twist nobody sees coming: the most advanced AI interfaces will look like a 1990s word processor. Bold, strikethrough, comment—these are the building blocks of a new human-AI language. Cleaner than chat, faster than code, more precise than both.
Neutrality is death. I’m taking a side: conversational UIs are a dead end for complex, multi-step agent tasks. They flatten reasoning into linear text, forcing you to rebuild context with every message. The answer isn’t a bigger context window. It’s a spatial canvas where you can see the whole plan, point, and edit.
Does this mean we abandon chat? Not entirely. Chat is fine for simple Q&A, brainstorming, or drafting emails. But when you’re debugging an agent’s reasoning, you need levels—a visual hierarchy that lets you drill into a step, annotate, and return. Chat gives you a tube. Direct manipulation gives you a map.
So next time you find yourself typing “No, not that, the other thing, the one I mentioned two messages ago,” stop. Open a local annotation UI. Draw a line through the nonsense. Write a comment. And watch how fast the AI finally understands.
The best way to talk to AI is not to talk at all. It’s to print out its plan, grab a red pen, and start crossing things out. That’s not a step backward. That’s the future.
FAQ
Q: Isn't conversational AI the whole point of LLMs? Why would we go back to point-and-click?
A: Chat is great for simple tasks, but for complex, multi-step agent plans, it's a cognitive nightmare. You lose context, you repeat yourself, and you can't visually edit the AI's reasoning. Direct manipulation—like strikethrough and comments—is faster and more precise for editing cognition.
Q: Can I use this approach with any AI agent, or is it only for specific tools?
A: The principle applies to any agent that generates a plan. The open-plan-annotator is a local UI that works with any agent's output. You can build your own or use similar patterns. The key is to move from linear chat to spatial editing.
Q: Isn't this just a step backward to old-school GUIs? Why not make chat smarter?
A: Making chat smarter (bigger context, better memory) is an arms race. The real breakthrough is offloading the verification burden from the user's working memory to a visual interface. It's not backward—it's ergonomic. Your brain is wired for spatial manipulation, not serial text.