When Everyone Has AI, Judgment Is the Real Differentiator

In July 2026, a service named Slopfix sparked discussion on Hacker News. Three senior engineers specialize in tidying up already running, but increasingly slow Vibe Coding projects. They concentrate on fixing them in a week, charging 10,000 dollars for full completion. They themselves use Claude Code, yet their website says one direct line: the agent doesn't get a vote.

A piece of seemingly absurd AI news

Slopfix's website states that the team consists of three senior engineers specializing in tidying up projects that are already running but become increasingly slow and more prone to cascading issues when new features are added.

They first offer free analysis of the codebase. If they judge there is potential for improvement, they agree on a code reduction target with the client and then concentrate on tidying it up in a week. The website gives an example of reducing 100,000 lines to 35,000 lines while maintaining the original functionality. They charge 10,000 dollars for full completion, and charge proportionally if only part of the target is met.

More interestingly, they themselves use Claude Code. The line on their website, 'the agent doesn't get a vote,' can be understood in Chinese as: Agents can participate in the work, but the decision-making power remains with the engineers.

Slopfix first has people review each scene and API endpoint, write a list of functionalities, and then organize repeated logic, replace unnecessary self-made frameworks, or rebuild parts that are already difficult to maintain. At delivery, they also leave a QA list, Claude.md, lint rules, and CI checks, reducing the probability of code regressing again.

Code will expand, this is not just a news headline.

The SlopCodeBench was publicly released in March 2026, studying how the quality of Coding Agent changes when it repeatedly expands its previously written code.

The current arXiv version of the experiment includes 36 problems, 196 checkpoints, 15 models, and 6 vendors. Researchers only give external functional requirements, letting the Agent decide its internal architecture, and then gradually add new specifications.

The results show that no tested configuration can fully solve all stages of a problem. Structural erosion occurs in 77% of the trajectories, and code length increases in 75.5% of the trajectories. Compared to 473 Python open-source projects, the code generated by the Agent is on average more lengthy and has higher structural erosion.

It still points out an important phenomenon: code that runs fine at this step does not necessarily mean it is easy to modify in the future. Prompts can improve the initial quality; however, the study does not prevent degradation after repeated iterations.

Where will the features return when tools become popular?

I have been talking about this for a while: when everyone can use AI, what is my unique feature?

You can use ChatGPT, and others can too. You can use Claude, Codex, and others can too. Once tools become popular, simply 'using them' will no longer be a differentiator.

What will truly pull the gap apart is returning to a few things:

1. Can you understand the actual problem you are trying to handle?
2. Do you know which limitations cannot be touched?
3. Do you have the ability to judge whether the result is usable or just looks complete?
4. How deep is your understanding of this industry, this job, and these users?

The same AI, handed to different people, can produce very different values. The difference usually comes from whether the user can provide enough professional context, whether they can discover what the AI missed, and whether they can continue to question even when the result seems reasonable.

Slopfix's approach makes this very specific. AI can find repetition, modify code, and run tests. Engineers need to first define functionality, draw the boundaries that cannot be broken, decide what is worth deleting, and finally take responsibility for the result.

This logic also applies to work outside of programming. Consultants need to judge whether advice fits the situation, instructors need to judge whether content suits the class of students, and managers need to judge whether a plan will affect existing processes. AI can help generate options, but professional workers still need to decide which option suits the current context.

Where to place AI

I will divide work into three layers simply.

First layer: Define the problem

What problem are you dealing with? Who is affected? What state do you hope to truly change after completion?

Second layer: Set judgment criteria

Which conditions cannot be broken? Which information needs to be verified? What situations count as completion? Who is responsible for the final result?

Third layer: Let AI accelerate execution

Organizing data, proposing options, finding repetition, generating drafts, testing, and checking can all be handed over to AI. The clearer the first two layers, the easier it is to turn the third layer's speed into real value.

Before handing over to AI, answer these three questions

If you've read this article and only want to do one thing, you can start with this list. Each time you hand over important work to AI, write down three answers:

1. What problem do I want to solve? First say the target, context, and the desired change in results.
2. What limitations cannot be touched? List facts, regulations, brand tone, budget, time, existing functions, or other red lines.
3. How do I judge whether the result is usable? Write down observable acceptance criteria to avoid ending with just 'it looks good'.

Then check again: What decisions did AI make during the process? Are there any parts that I need to personally take back and judge?

You can also directly put the following prompt into the workflow:

Make your expertise deeper, then let AI amplify it

After AI becomes widespread, operating tools will still be important, but it is becoming more like a basic ability.

What is more worth investing in next is to gain a deeper understanding of a particular industry, a group of people, or a type of problem. Go to understand how things operate on the ground, why common mistakes happen, which details affect the outcome, and what costs cannot be tolerated. Then organize these experiences into judgment criteria, checklists, cases, and workflows, and hand them over to AI to assist in execution.

Using AI in this way can steadily amplify your expertise. When there is a lack of problem definition, limitations, and acceptance criteria, AI may also simultaneously amplify the unnoticed chaos within the work.

Tools will keep changing. What is truly worth accumulating is your ability to understand problems, your criteria for making judgments, and your long-term understanding of a field.

Source and Verification Boundary

• Slopfix Official Website: Team, process, pricing, reduction examples, Claude Code usage methods, and delivery content are all self-reported by the service provider.
• Slopfix Discussion Thread on Hacker News: Can view public discussions and doubts about the service after its release.
• SlopCodeBench Paper: This article uses the arXiv version data from 2026-07-17. The paper may still be updated, so confirm the version again when citing.
• Tom's Hardware Report: Used to cross-verify news background. Key service information should still be based on the official website.