---
source: https://iinventedlight.com/library/notes/ao-rules-instruction-adherence.md
title: "Does rules-file structure change whether agents follow it?"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2605.10039] Instruction Adherence in Coding Agent Configuration Files: A Factorial Study of Four File-Structure Variables, CC BY 4.0; arXiv, [2607.27250] Do Context Files Help Coding Agents? A Two-Agent Ablation Study on Real Repositories, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was rerun; only the fetched arXiv abstract pages were read. Results are as reported by the authors."
category: agent-operations
topic_area: "agent operations: rules files"
tags: agents, coding-agents, instructions, compliance
sources:
  - id: S1
    title: "[2605.10039] Instruction Adherence in Coding Agent Configuration Files: A Factorial Study of Four File-Structure Variables"
    publisher: arXiv
    url: https://arxiv.org/abs/2605.10039
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2607.27250] Do Context Files Help Coding Agents? A Two-Agent Ablation Study on Real Repositories"
    publisher: arXiv
    url: https://arxiv.org/abs/2607.27250
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Does rules-file structure change whether agents follow it?

## Short answer
In one factorial study, file size, position, architecture and conflicts had no detectable effect on compliance. What mattered was session length: compliance fell as the agent generated more functions. [S1] A separate ablation also found context strategy did not move correctness. [S2]

## Practice
- Do not over-tune file layout.
- Re-check key rules during long sessions, or split work into shorter ones.

## Claims
- id: C1
  claim: Structural variables showed no detectable contrast.
  source: S1
  evidence: "None of the four structural variables or three two-way interactions produces a detectable contrast"
- id: C2
  claim: Each extra function lowered compliance odds about 5.6%.
  source: S1
  evidence: "approximately 5.6% lower odds of compliance per step"
- id: C3
  claim: 1,650 sessions were measured.
  source: S1
  evidence: "across 1,650 Claude Code CLI sessions"
- id: C4
  claim: 288 runs were evaluated.
  source: S2
  evidence: "288 evaluated runs with gold-test evaluation"

## Sources
- [S1] [2605.10039] Instruction Adherence in Coding Agent Configuration Files: A Factorial Study of Four File-Structure Variables. arXiv. https://arxiv.org/abs/2605.10039 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2607.27250] Do Context Files Help Coding Agents? A Two-Agent Ablation Study on Real Repositories. arXiv. https://arxiv.org/abs/2607.27250 . License: CC BY 4.0. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).