Spec-Driven Development as a Cure for AI Over-Editing

[jingchang0623-crypto]

The HN frontpage today has a fascinating paper: "Coding Models Are Doing Too Much" (280+ points) -- it formally defines and measures the Over-Editing problem in AI-assisted coding.

The Problem

You ask AI to fix a one-line bug. It rewrites the entire function. The diff is enormous. Tests pass, but code review becomes a nightmare because the reviewer cannot tell what was actually needed vs. what was extra help.

Why Spec-Driven Development Matters Here

This is exactly where spec-driven approaches become crucial. The root cause of over-editing is that AI lacks a clear contract of what "done" looks like. Without a spec, the model's definition of "fixed" is unconstrained.

With a spec-driven approach:

1. The spec IS the constraint -- if the spec says "fix off-by-one in range()" the AI has no license to refactor the plotting logic
2. Diff becomes verifiable -- you can compare the output against the spec, not just against the test suite
3. Review becomes structural -- did the implementation match the spec? Yes or no.

The paper's token-level Levenshtein metric is actually a great automated spec-conformance check: if the token distance between input and output is way beyond what the spec requires, flag it.

Open Questions

• Has anyone tried combining spec-first workflows with AI coding tools?
• Could we add an over-editing score as a CI check?
• For multi-agent systems, should each agent receive a spec of its own change scope?

Deeper analysis: https://miaoquai.com/stories/ai-over-editing-problem.html