{"paper":{"title":"Guardrails Beat Guidance: A Large-Scale Study of Rules, Skills, and Persistent Configuration for Coding Agents","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Negative constraints raise AI coding agent success rates while positive instructions lower them, and random rules match expert ones.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Bing Zhu, Guanghui Wang, Peiyang He, Wei Qiu, Xing Zhang, Yanwei Cui, Ziyuan Li","submitted_at":"2026-04-13T07:10:01Z","abstract_excerpt":"Random rules improve a coding agent's task performance as much as expert-curated ones (both $+13.8$pp on a discriminative subset of SWE-bench Verified), and in our data every individually beneficial rule is a negative constraint (\"do not refactor unrelated code\"), while every individually harmful one is a positive directive (\"follow code style\"). We arrive at these findings through the first large-scale controlled study of agent rule files (\\texttt{CLAUDE.md}, \\texttt{.cursorrules}, and the broader family of agent skills, plugin manifests, and persona definitions): we scrape 679 rule files (25"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Rules improve performance by 7--14 percentage points, but random rules help as much as expert-curated ones -- suggesting rules work through context priming rather than specific instruction. Negative constraints are the only individually beneficial rule type, while positive directives actively hurt.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That differences in agent success rates are caused by the rule types themselves rather than unmeasured factors in how rules are inserted into prompts or variations across agent runs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Negative constraints in agent rules improve coding performance via context priming while positive directives degrade it, with collective rules remaining helpful up to 50 rules.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Negative constraints raise AI coding agent success rates while positive instructions lower them, and random rules match expert ones.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e7dd00151feaf1be5a651358c4b6a755fc576a6d4ca45b0f75e2f20f231d11e7"},"source":{"id":"2604.11088","kind":"arxiv","version":2},"verdict":{"id":"a2e26d41-360e-4016-8608-abd8319657b3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T15:33:36.613027Z","strongest_claim":"Rules improve performance by 7--14 percentage points, but random rules help as much as expert-curated ones -- suggesting rules work through context priming rather than specific instruction. Negative constraints are the only individually beneficial rule type, while positive directives actively hurt.","one_line_summary":"Negative constraints in agent rules improve coding performance via context priming while positive directives degrade it, with collective rules remaining helpful up to 50 rules.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That differences in agent success rates are caused by the rule types themselves rather than unmeasured factors in how rules are inserted into prompts or variations across agent runs.","pith_extraction_headline":"Negative constraints raise AI coding agent success rates while positive instructions lower them, and random rules match expert ones."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.11088/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}