Last year, I hit a wall: I simply could not get my AI assistants to follow my instructions to edit my files accurately and reliably. I found this to be true regardless which model or client I chose, no matter how well I documented, and across all file types.

I began focusing on a specific scenario: The agent echoes back a perfectly reasonable plan to revise a file, they call their tools, and they announce completion; but, the file is corrupted. Even if not literally broken, the diff shows wholesale replacements, many unrelated to the underlying issue, when small, surgical modifications were warranted. I call this last-mile failure pattern "Execution Slop".

After investigation, I concluded that Execution Slop cannot be fixed through prompt engineering or by paying for more expensive tokens, because the AI file-editing tools themselves are broken. All major AI coding assistants use the same string-replacement strategy for editing under the hood. Agents can't visualize their changes before committing them to disk, get no warning that they're about to break something, can't roll back changes atomically if they realize they made a mistake, and often can't even insert or delete at a specific line or line range (let alone a particular column), without also echoing everything around it.

So I spent nearly a year building something completely different. HIC Mouse gives AI agents line- and coordinate-based editing through a natural syntax that allows agents to edit concisely by declaring region boundaries instead of forcing them to use string replacement. All multi-operation and large operations are automatically staged in memory before touching disk, triggering a Dialog Box mode, in which the agent can save, cancel, inspect, or refine. If something goes wrong, the agent can roll back edits atomically. If most of a batch succeeds but one operation fails, the agent can fix just the failure without discarding the rest. And agents are given embedded contextual guidance at every tool call.

To validate rigorously that HIC Mouse genuinely improves outcomes, I ran three preregistered confirmatory studies (N=67 paired runs) comparing Mouse-enabled AI assistants running in isolated Docker containers performing timed, realistic file-editing tasks ranging in difficulty, against identically configured agents using built-in editing tools. I've uploaded the technical report and statistical analysis with all the details, but the bottom line is that Mouse dramatically improved performance (Cohen's h > 2 or "massive" effect size on multiple metrics), across every dimension that I studied -- capability, speed, cost, reliability, and most importantly, accuracy.

We have now officially launched, and HIC Mouse is available for download through the VS Code Marketplace and Open VSX. Mouse works with VS Code, Cursor, and Kiro, and it's compatible with GitHub Copilot, Claude Code, and other MCP clients.

Please consider installing HIC Mouse, and let me know what you think! I really hope that it genuinely makes a positive difference for you.

Now let me ask you: How often do you encounter Execution Slop, and what are you doing to avoid it?

Comments URL: https://news.ycombinator.com/item?id=48036814

Points: 2

# Comments: 1

A great article with some really beautiful examples of zugzwang is: https://www.chesshistory.com/winter/extra/zugzwang.html. There's a very nice discussion at the end as well of a disagreement along just these lines as to what truly constitutes zugzwang, between Hooper and Myers.