We charge $10k a week to delete AI-generated code

Original Article ↗ Hacker News Discussion ↗

Business model: “Slopfix” refactoring service

  • Service: one-week engagements, three senior engineers, ~$10k, focused on deleting and refactoring AI-generated “slop” code.
  • They commit to a code-size reduction target; payment is proportional to how much reduction they achieve.
  • Heavy use of LLMs as tools (e.g., Claude Code) for large-scale refactors, but humans drive architecture and decisions.
  • Two-week “warranty” window drew criticism as too short; provider claims in practice they don’t abandon clients and bugs are easier to fix now.

Is this actually new?

  • One camp: this is just the latest version of cleaning up legacy “big ball of mud” systems, outsourced code, cloud/crypto/whatever hype fallout.
  • Other camp: AI has created an unprecedented explosion of LOC; AI-slop will dwarf classic legacy tech and create long-term “slopfield” work.

Quality of AI-generated code

  • Experiences vary widely:
    • Some report AI-assisted code that’s cleaner, better documented, and easier to maintain than pre-AI code when used by skilled engineers.
    • Others see extreme “vibe-coded” mess from non-engineers or inexperienced users: brittle, entangled architectures, dangerous shortcuts (e.g., destructive DB migrations).
  • Several posters distinguish:
    1. Pure vibe coding by non-devs.
    2. Non-coders with process awareness using AI.
    3. Engineers using AI with strong review and structure.
    • Moving clients from (1) to (3) is seen as highly valuable.

Using AI to clean AI

  • Skeptics compare “AI in, AI out” to repeated lossy transcoding: errors compound rather than cancel.
  • Supporters say AI is a power tool: in the hands of experienced devs, it accelerates targeted refactors, code profiling, and pattern-based cleanup.
  • Common pattern: write or improve tests (e2e, unit, mutation), define strict duplication/linting rules, then iteratively cut and refactor with AI under CI.

Feasibility, scope, and risk

  • Doubts about understanding complex business rules in a week and catching regressions, especially with weak test suites.
  • Proponents claim:
    • Much AI-slop is early-stage and not yet burdened by a decade of corner cases.
    • Work can be decomposed into “conceptually hard” human tasks plus “repetitive pattern” tasks delegated to AI.
    • Initial scans, client walkthroughs, and architecture decisions are human-led.

Market, incentives, and long-term concerns

  • Target customers are often non-technical founders who overused AI and now face unmaintainable repos.
  • Some predict a perpetual cycle: clients refactor, then re-vibecode on top, generating recurring consulting work.
  • Others argue the “real” fix is human-led rewrites, but many push back that wholesale rewrites are rarely optimal.
  • Concern raised about industry dependence on proprietary LLMs (Anthropic/OpenAI) as core infrastructure and the risk of future pricing or availability shocks.