[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-78183":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":16,"stars30d":17,"stars90d":15,"forks30d":15,"starsTrendScore":18,"compositeScore":19,"rankGlobal":10,"rankLanguage":10,"license":20,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":21,"hasPages":21,"topics":23,"createdAt":10,"pushedAt":10,"updatedAt":34,"readmeContent":35,"aiSummary":36,"trendingCount":15,"starSnapshotCount":15,"syncStatus":14,"lastSyncTime":37,"discoverSource":38},78183,"meta-llm-charter","entropyvortex\u002Fmeta-llm-charter","entropyvortex","Compact operational charter that turns LLM coding agents into disciplined principal engineers. Eleven rules + one meta-rule.","",null,"TypeScript",241,19,2,0,3,140,1,3.9,"MIT License",false,"main",[24,25,26,27,28,29,30,31,32,33],"ai","ai-agents","claude","coding-agents","constitution","cursor","frontier-ai","llm","llm-agents","prompt-engineering","2026-06-12 02:03:46","# META v2.0 — LLM Agent Engineering Charter  \n**(with Zero-Pause Native Execution Layer)**\n\n**One file. Eleven rules + continuous-execution layer. One meta-rule. One bias.**\n\nA compact, operational constitution that turns frontier coding agents (Claude Code, Cursor, etc.) from eager-junior behavior into disciplined principal-engineer execution **with unbroken velocity**.\n\n### Quickstart\n\n```bash\n# Drop the charter into your project root\ncurl -O https:\u002F\u002Fraw.githubusercontent.com\u002Fentropyvortex\u002Fmeta-llm-charter\u002Fmain\u002FCLAUDE.md\n```\n\n- **Claude Code**: Reads `CLAUDE.md` automatically.\n- **Cursor**: Paste contents into Cursor Rules (or `.cursor\u002Frules`).\n- **Other agents**: Use as high-priority system prompt.\n\n[View raw CLAUDE.md](https:\u002F\u002Fraw.githubusercontent.com\u002Fentropyvortex\u002Fmeta-llm-charter\u002Fmain\u002FCLAUDE.md)\n\n### What’s new in v2.0\n**Zero-Pause Native Execution Layer** is now baked directly into `CLAUDE.md`.  \nAny task that mentions “Zero-Pause”, “zero pause”, “ZP-”, or the activation phrase instantly enables:\n- Continuous forward momentum (no artificial phases, no mid-task questions)\n- `humanpending.md` protocol for true human-gated items\n- Parallel ASI orchestration (minimum 7 specialized threads)\n- Zero session-size anxiety\n\nThe original META v1.3 rules (R1–R11) remain untouched and in force at all times.\n\n### Use as Grok Skill on grok.x.ai (web \u002F mobile)\n\nGrok now supports **Custom Instructions** and named **Skills**.\n\n→ See [GROK-META.md](GROK-META.md) for the one-click setup (Custom Instructions recommended — works instantly on every chat).\n\n### Why this exists\n\nLLM coding agents are incredibly capable but consistently fail in the same senior-level ways. META closes those gaps; Zero-Pause closes the velocity gaps.\n\n### Core Philosophy\n\n**Bias — Earned Conservatism**  \nDefault to first-principles rigor. Quality dominates token count. Move boldly on local, reversible, test-covered work. Apply explicit, named caution only on high blast-radius or low-reversibility moves.\n\n**META-0 — Situated Judgment Overrides Rules**  \nThese rules are scaffolding. When first-principles analysis of the actual situation conflicts with a rule, follow the analysis. Name the override, justify it, and be evaluated on judgment quality + ground-truth outcomes — not rule compliance.\n\nThe eleven rules (R1–R11) + Zero-Pause layer (ZPR1–ZPR4) operationalize decomposition, decisiveness, verification, scope control, epistemic tagging, pushback, reversibility, **and relentless continuous execution**. Full charter is in [CLAUDE.md](CLAUDE.md).\n\n### What the charter actually changes\n\n- **R5 + R8**: Forces reproduction before repair and tags every claim (`executed` \u002F `inspected` \u002F `assumed`).\n- **R9**: One clear, evidence-based pushback on bad premises — then defer and document dissent.\n- **R4 + R10**: Bounded refactoring and reversibility-weighted boldness.\n- **Zero-Pause layer**: Unbroken execution, pre-work questions only, parallel orchestration, and `humanpending.md` handling.\n\n### Evaluation\n\nThe repo includes a reproducible TypeScript + Docker A\u002FB test harness in [`evals\u002F`](evals\u002F). It runs agents against five synthetic fixtures engineered to trigger classic agent failure modes.\n\n**Latest smoke-test results (May 12, 2026)**:  \nCharter variant won outright on 3\u002F5 tasks and tied on 2\u002F5 against a generic “principal engineer” baseline. Full details, raw CSVs, and judge transcripts are in the evals directory.\n\n(The harness is public and cheap to run: `cd evals && npm run smoke`.)\n\n### Known limitations\n\n- Still early (v2.0, single-author origin).\n- Performance varies by base model — strongest with frontier Claude\u002FSonnet-class models.\n- Can produce over-caution on fuzzy\u002Fcreative\u002Fexploratory work (Zero-Pause helps here).\n- Not magic: extremely ambiguous requirements can still overwhelm any system prompt.\n\n### When to use META\n\n**Best for**  \nSerious software engineering where correctness, maintainability, long-term system health, **and velocity** matter.\n\n**Less ideal for**  \nPure exploration, rapid UI prototyping, research spikes, or contexts where you explicitly want maximum speed over discipline (though Zero-Pause narrows this gap significantly).\n\n### Contributing\n\nMost valuable contributions right now:\n1. Running the eval harness on new models\n2. High-quality held-out fixtures (especially adversarial Zero-Pause cases)\n3. Sanitized real-world case studies\n\nSee [CONTRIBUTING.md](CONTRIBUTING.md) for details.\n\n### Lineage\n\nBuilt on the foundational minimal principles from  \n[forrestchang\u002Fandrej-karpathy-skills](https:\u002F\u002Fgithub.com\u002Fforrestchang\u002Fandrej-karpathy-skills).\n\n### License\n\nMIT\n\n---\n\nBy [entropyvortex](https:\u002F\u002Fgithub.com\u002Fentropyvortex).  \n\nFeedback, evals, and war stories welcome.","该项目提供了一套简洁的操作章程，旨在将LLM编码代理转变为具有纪律性的首席工程师。核心功能包括11条规则和一个持续执行层，以及一条元规则，通过这些规则确保编码代理在执行任务时保持高效且有序。技术特点上，项目采用TypeScript编写，并引入了“零暂停”原生执行层，以保证任务的连续性和无间断推进。适用于需要提升AI编码代理如Claude Code、Cursor等在复杂开发环境中的表现力和效率的场景，尤其适合追求代码质量和执行速度平衡的高级开发团队。","2026-06-11 03:56:33","CREATED_QUERY"]