Recoleta
GitHub
Source note

ZORO: Active Rules for Reliable Vibe Coding

Authors
Jenny Ma, Sitong Wang, Joshua H. Kung, Lydia B. Chilton
Source
arXiv
Published
Collection
Software Intelligence
Code IntelligenceHuman AI InteractionSoftware Foundation ModelAutomated Software ProductionMulti Agent Software Engineering
Open arXivMarkdown source

Summary

Zoro turns rules files such as AGENTS.md from passive text into active controls during AI-assisted coding. It ties rules to plan steps, blocks progress until required rules are proven, and lets users revise rules from concrete failures.

Problem

  • In vibe coding, rules files are common, but agents often stop following them as a session goes on because rules are only loaded as static context.
  • Developers cannot easily see which rules were applied, whether they were followed, or how to improve weak or outdated rules.
  • This matters because teams use these rules to enforce architecture, workflow, UI, and safety constraints; if the agent skips them, developers must manually inspect, correct, and repeat instructions.

Approach

  • Zoro adds an Enrich-Enforce-Evolve workflow on top of existing coding agents such as Codex, Claude Code, Cline, and Cursor.
  • Enrich: after the agent creates an initial plan, Zoro matches rules to specific plan steps and can split or rewrite steps so the rules are attached where they matter.
  • Enforce: during execution, the agent must call Zoro CLI commands to mark step progress and submit proof that each required rule was followed; for testable rules, it must also provide unit tests before moving on.
  • Evolve: users review the rule evidence in the UI, add in-situ notes when a rule was applied poorly, and Zoro aggregates those notes to refine the ruleset with LLM help.
  • The system is agent-agnostic through a ZORO.md instruction file and a shared .zoro directory that stores the enriched plan, rule metadata, evidence, and user notes.

Results

  • The paper reports a 57% increase in rule following with Zoro compared with standard vibe coding.
  • The technical evaluation covers 36 vibe coding sessions and says Zoro improves rule adherence while maintaining feature-completion ability relative to a standard coding agent.
  • The user study includes 12 participants and reports a shift in how people control agents: users moved from prompt engineering toward rule engineering.
  • The formative design work included 10 programmer interviews plus observation of 3 programmers in sessions lasting 2-3 hours each.
  • The excerpt does not provide a metric definition, dataset name, variance, or exact baseline score beyond the 57% relative improvement claim.

Link