A set of four project-level CLAUDE.md instructions attributed to Andrej Karpathy tells Claude Code to surface assumptions, avoid unrequested abstractions, make only necessary edits, and define clear success criteria. Community posts and a GitHub adaptation describe the rules as a lightweight harness for recurring coding-agent failure modes, with some reports claiming substantial error reductions and viral adoption. The supplied results do not include Karpathy’s original post or reliable validation of the performance and adoption claims, and one snippet cautions that the rules may not cover multi-step agent loops.
Scott already holds this position in “CLAUDE.md Pattern” and “Your AI Can Code. You Just Don’t Know How to Drive It.”: slim persistent instructions can encode lessons, surface assumptions, constrain scope, and discourage premature abstraction. The emerging four-rule packaging is a publishing opportunity around his broader doctrine, but he would need to distinguish lightweight policy from the plan/test/verification loops he considers necessary for reliable delivery; the supplied adoption and performance claims remain unvalidated.
ip:concept.claude-md-patternip:source.your-ai-can-code-you-just-don-t-know-how-to-drive-it-ebookip:concept.fat-agents-md-anti-patternip:concept.plan-mode-disciplineip:concept.test-first-agent-workflowradar:concept.coding-agentsradar:concept.agent-harnessradar:concept.claude-coderadar:karpathy-llm-wiki-adoption
queries asked of Scott's wikis
- coding-agent harness constraints and guardrails
- minimal diffs and unrelated-edit prevention
- assumption surfacing before agent execution
- project instruction files as persistent agent policy
- anti-abstraction rules for generated code
- single-shot coding rules versus multi-step agent loops
2026-07-24T14:29:23Z
The added HN discussion reinforces the shift toward closed-loop validation and broader scoped-instruction harnesses, not adoption of Karpathy’s four-rule bundle. After repeated independent examples of the broader doctrine but none of the recognizable package, the specific episode is superseded by the more general CLAUDE.md harness pattern.
2026-07-24T14:21:24Z
evidence attached: hn.story.49033388 — shared external link with case evidence
2026-07-24T13:23:18Z
The apparent trigger adds no package-level adoption beyond evidence already assessed; repeated adjacent CLAUDE.md implementations corroborate the broader harness doctrine, not spread of Karpathy’s specific four-rule bundle. Treat further engagement-only updates as repetitive amplification and wait for a direct citation, implementation, or validation signal.
2026-07-24T12:26:51Z
The HN monorepo example is another adjacent implementation of scoped-instruction harness design, not evidence of Karpathy's specific four-rule bundle spreading. After twenty-plus repricings, the pattern is stable: the broader CLAUDE.md/persistent-instruction doctrine keeps getting corroborated while the package-level hypothesis stays unvalidated. Cooling review cadence further absent a genuine adoption or citation signal.
2026-07-24T12:21:10Z
evidence attached: hn.story.49034424 — Its path-scoped rules, hooks, and sub-agent structure provide a concrete coding-harness example relevant to the emerging CLAUDE.md pattern.
2026-07-24T11:26:26Z
The new examples strengthen the already-established broader pattern of repository guidance files, while the trust critique reinforces that static Markdown rules need closed-loop validation. Neither provides independent adoption of Karpathy’s specific four-rule bundle, so the package-level spread hypothesis remains uncorroborated.
2026-07-24T11:21:01Z
evidence attached: reddit.post.1v57vm8 — A concrete CLAUDE.md design-system workflow supports the case that repository guidance files are spreading as a coding-agent harness pattern.
2026-07-24T11:21:01Z
evidence attached: reddit.post.1v58gax — The article directly bears on whether Markdown specs and skills are becoming a reliable coding-agent harness pattern, while arguing that closed-loop validation is needed.
2026-07-24T08:24:40Z
The nominal evidence trigger adds no independent adoption of Karpathy’s specific four-rule package; unchanged adjacent examples continue to corroborate only the broader CLAUDE.md harness doctrine. Further engagement-only reobservations are repetitive amplification, so defer review until a package-level implementation, citation, or validation appears.
2026-07-24T00:21:13Z
The nominal evidence trigger contains no substantive new package-level adoption; it is another reobservation of adjacent CLAUDE.md practices. Karpathy’s four-rule bundle remains a single viral adaptation with mixed practitioner results, so the specific spread hypothesis stays uncorroborated.
2026-07-23T23:29:43Z
No new evidence attached this cycle beyond prior repricing; engagement continues to be repetitive amplification of adjacent CLAUDE.md scope-control practices, still without independent adoption of Karpathy's specific four-rule package. Case remains a cooled watch pending a genuine implementation or citation signal.
2026-07-23T21:29:01Z
The attached implementation remains evidence for the broader scope-and-verification harness doctrine, not independent spread of Karpathy’s four-rule bundle. With package-level adoption still absent, further adjacent examples and engagement are repetitive amplification rather than advancement.
2026-07-23T20:29:19Z
No new package-level adoption or validation changes the case: the record still corroborates only the broader CLAUDE.md harness doctrine, while Karpathy’s specific four-rule bundle remains a single viral adaptation with known guardrail problems. Further engagement-only reobservations are repetitive amplification and should not trigger frequent review.
2026-07-23T19:29:08Z
The new implementation further supports project-level verification and scope-control instructions, but remains adjacent to rather than adoption of Karpathy’s recognizable four-rule package. The case still rests on one viral adaptation with repeated amplification, while practitioner feedback shows the rules require contextual guardrails.
2026-07-23T19:21:21Z
evidence attached: reddit.post.1v4nqeh — Independent coding-agent use reinforces CLAUDE.md rules that require screenshot checks to constrain unrelated or incorrect UI edits.
2026-07-21T23:31:15Z
The attached record still adds no independent adoption of Karpathy’s specific four-rule package; it continues to corroborate only the broader, already-known persistent-instruction pattern. Repetitive engagement and adjacent practices no longer justify frequent checks without a package-level implementation, citation, or validation signal.
2026-07-21T22:24:37Z
The apparent velocity spike is only six additional votes with no new discussion, implementation, or independent adoption of Karpathy’s specific four-rule package. Repeated engagement noise continues to support neither recognizable spread nor resolution, so the case remains a cooled watch.
2026-07-21T18:28:16Z
The latest reobservation adds no independent evidence that Karpathy’s four-rule package is spreading as a recognizable pattern; it remains a popular seed surrounded by broader, already-known CLAUDE.md practices. Repeated amplification without package-level implementations or validation does not advance the case.
2026-07-21T16:33:56Z
No substantive new evidence shows Karpathy’s specific four-rule package spreading independently; the record still supports only the broader persistent-instruction harness pattern. Repeated reobservation without package-level adoption or validation adds no meaning and keeps the case cool.
2026-07-21T15:31:52Z
The negligible engagement increase adds no independent adoption signal; evidence still supports the broader CLAUDE.md harness doctrine rather than spread of Karpathy’s specific four-rule package. Keep watching, but the case remains cooled and uncorroborated at the package level.
2026-07-21T13:23:19Z
The new workflow report independently reinforces explicit scope-control instructions as a useful harness technique, but it does not show Karpathy’s four-rule package spreading as a recognizable pattern. Evidence continues to support the broader doctrine rather than the case’s specific adoption hypothesis.
2026-07-21T13:21:08Z
evidence attached: reddit.post.1v2hxa1 — A firsthand coding-agent workflow reports that explicit scope and out-of-scope handling instructions prevent silent omissions, providing contextual evidence for harness instruction patterns.
2026-07-20T23:21:17Z
The added practitioner example supports CLAUDE.md as a general harness convention but does not independently show Karpathy’s four-rule package spreading as a recognizable pattern. With no fresh engagement or validation—and substantive reports that the rules can induce shortcuts—the case merits monitoring but has cooled.
2026-07-20T23:21:01Z
evidence attached: reddit.post.1v1zzga — Independent practitioner adoption of parent-level CLAUDE.md instructions supports the broader spread of persistent instruction files as a coding-agent harness pattern.
2026-07-20T06:10:52Z
grounded: known/medium — Scott already holds this position in “CLAUDE.md Pattern” and “Your AI Can Code. You Just Don’t Know How to Drive It.”: slim persistent instructions can encode l
2026-07-20T06:08:40Z
case created — A highly engaged adaptation provides an initial signal that builders are operationalizing the rules rather than merely quoting them.