A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
| Rating | Summary | |
|---|---|---|
| Quality | decent (13/24) | Massively adopted and well-documented but the repo infrastructure is nearly bare — no license, no releases, no CI, zero open issues. |
| PAI Relevance | watch (0.63) | Language-agnostic and trivially drop-in; the four principles could sharpen PAI's own CLAUDE.md, but partial overlap with existing agent behavior configuration keeps it from crossing the integrate threshold. |
13/24 — dormant-or-abandoned / adequately-documented / solid
Failed:
Passed:
Failed:
Passed:
Failed:
Passed:
| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 1 | The "Goal-Driven Execution" principle — transform imperative tasks into declarative success criteria for autonomous looping — maps usefully to PAI's Loop and Evals skills, but is not architecturally novel enough to score 2. |
| Integration Readiness | 2 | Pure markdown; language-agnostic; drop-in via curl or copy-paste merge into PAI's existing CLAUDE.md. No adapter code, no runtime dependency. |
| Overlap Risk | 1 | Partial overlap with PAI's existing agent behavior configuration (the ~27 agents each carry instruction context), but no dedicated "coding behavior principles" skill exists by name in the Capability Manifest. |
| Gap Fill | 1 | PAI's Capability Manifest includes agent orchestration and eval tooling but no explicit coding-session discipline layer; the four principles address a limited functional area PAI could benefit from without it being a critical gap. |
Composite: 0.63
Before starting a multi-file refactor with Claude Code: Drop the repo's CLAUDE.md verbatim into the project root and run one scoped task — a single function rewrite or a bug fix with a clear acceptance condition. Compare Claude's output behavior (does it ask clarifying questions before writing? does it stay within the stated scope?) against baseline sessions without the file. This gives a concrete signal on whether the four principles meaningfully change output quality before investing in per-project customization.
When the four-principle pattern starts appearing in Claude Code's official documentation or system prompt defaults: The 158K-star adoption is a strong signal Anthropic is tracking this. Monitor the Claude Code changelog and official CLAUDE.md documentation for first-class support (typed rules, conditional application, task-scoped overrides). When that lands, this community file becomes the baseline to benchmark any official implementation against — knowing its current principles makes that comparison immediately legible.
If you start accumulating per-project CLAUDE.md files: Revisit the fork network (16K forks) for principle extensions beyond the original four — particularly forks targeting frontend or TypeScript-heavy workflows where scope creep failure modes differ from general coding tasks. Only worth pulling from forks that show evidence of real usage (commit history, issues), not speculative additions.
Category: LLM & Prompt Tooling
In this category: mattpocock--evalite (the only prior taxonomy entry; focused on LLM evaluation tooling rather than behavior guidelines)
Standing: Thematically distinct from evalite — this is behavioral prompt engineering for coding sessions, not evaluation infrastructure — and sits closer in spirit to the "AI Coding Agent Skills" overlap cluster alongside mattpocock--skills and humanlayer--12-factor-agents.
Density: 6/10 — Available: full README, all repo metadata (stars, forks, dates, archived status, language, license), description, overlap cluster assignments from landscape. Missing: dependency manifest (none exists), CI configuration, actual CLAUDE.md file content (contents inferred from README descriptions only), release notes, contributor activity data.
The 158K-star count is anomalous for a single-file markdown repo created in January 2026 — it suggests a viral moment tied directly to Karpathy's original tweet rather than sustained organic growth. The health score (2/8) reflects a bare-bones repo infrastructure that is entirely appropriate for the artifact type; penalizing it for lacking releases or CI is correct by the probe rules but somewhat artificial. The standalone score of 13 reflects that the content is genuinely well-crafted and widely validated, even if the repo itself has no engineering substance beyond the markdown file.