Introducing AntiSlopScan: engineering judgment for AI-assisted code

#ai #architecture #opensource #softwaredevelopment

A pull request can pass its tests and still make the next change harder.

AI makes it easier to produce that pull request. We still need to ask whether the design makes sense, whether it repeats existing logic, and what happens when something fails.

I built AntiSlopScan to help developers investigate those questions. It is an open-source, self-hosted audit appliance for code quality, architecture and systems-thinking indicators. You can run it from the command line, use its web interface, or add it to CI.

Software factories need useful feedback

The software-factory idea interests me because it connects the whole development process: describing a problem, implementing a change, checking the result and learning from it.

AntiSlopScan contributes an inspection step to that process. It combines established scanners with independent structural analysis, then collects the evidence in one report. A developer can review duplication, complexity, dependency cycles and configured boundary violations alongside scanner coverage.

I want that feedback to help us correct a design while a change is still small.

Think beyond the function

Systems thinking asks us to examine relationships, feedback and consequences over time. Donella Meadows' work on leverage points is a useful starting point.

Consider a retry added to one service. It may make that service appear more reliable while increasing pressure on an already overloaded dependency. Understanding the outcome requires a view beyond the edited function.

A repository scan can reveal some structural evidence. It cannot establish every runtime feedback loop or business requirement. AntiSlopScan leaves unsupported assessments ungraded, and historical comparisons need a compatible baseline before they can show deterioration.

Make architectural intent checkable

A rule such as “domain code must not depend on transport code” gives a team something concrete to verify.

That connects to evolutionary architecture and fitness functions: make changes in small steps and check whether important architectural properties still hold. An executable boundary check gives reviewers useful evidence; they still decide whether the boundary is appropriate.

I see anti-slop practices as an emerging discipline for AI-assisted development, built on familiar engineering habits: understand the requirement, prefer a simple design, keep changes reviewable, test behavior and examine the effects across the system.

The talks that shaped this project

Why Coding Agents Keep Making Your Codebase Worse, with Dex Horthy on Beyond Coding, explores maintainability and practical agent workflows.
Horthy's Why Software Factories Fail challenges the assumption that passing tests settles long-term code quality.
Mario Zechner's Building pi in a World of Slop argues for keeping agent work understandable and retaining human judgment over consequential changes.

For further perspectives, try Vaibhav Gupta on automated verification and Addy Osmani on engineering judgment.

Try it on a repository

The Linux x86_64 alpha image can audit a read-only source mount and write reports to a separate folder:

docker run --rm --network none \
  -v /absolute/path/to/source:/workspace:ro \
  -v /absolute/path/to/reports:/reports \
  jfscoertzen/antislopscan:1.0.0-alpha.1 \
  cli --source /workspace --output /reports

Create the reports directory first and make it writable by UID 1000. You get HTML, PDF, JSON and SARIF reports. The README covers the web interface, local installation and configuration.

Core scanning works offline after installation. Optional LLM architecture review needs explicit source-transmission consent. The tool flags code patterns; it does not determine who wrote the code. Scores and findings are review aids, not proof that a system is correct.

Developers still own the decisions

I am excited about using agents to explore ideas and reduce repetitive work. I want the time we gain to improve how we understand users, explain trade-offs and design systems that other people can maintain.

Developers bring that understanding to a team. We can learn new tools while continuing to practice the work that makes software useful and dependable.

Try AntiSlopScan, inspect a finding, and open an issue with a reproducible example where the evidence or explanation needs improvement. It is an Apache-2.0 alpha project, and that kind of feedback will help shape it.

DEV Community