Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add Qiu-Dong88/super-nvwa --skill andrej-karpathy-perspectivegit clone --depth 1 https://github.com/Qiu-Dong88/super-nvwaWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/qiu-dong88/super-nvwa/andrej-karpathy-perspective)<a href="https://agentmods.dev/skills/qiu-dong88/super-nvwa/andrej-karpathy-perspective"><img src="https://agentmods.dev/badge/skills/qiu-dong88/super-nvwa/andrej-karpathy-perspective/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/qiu-dong88/super-nvwa/andrej-karpathy-perspective"><img src="https://agentmods.dev/badge/skills/qiu-dong88/super-nvwa/andrej-karpathy-perspective.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00247 | $0.08443 |
| Opus 5 | $0.00123 | $0.04222 |
| Sonnet 5 | $0.00049 | $0.01689 |
| Haiku 4.5 | $0.00025 | $0.00844 |
Grade A, and why
andrej-karpathy-perspective scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 11d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
This is a copy
88% identical to andrej-karpathy-perspective — 158 lines differ, which has more behind it and is treated as the original. This page carries a canonical link to it rather than competing with it.
How it starts
The opening of the file, as written. The whole thing — 508 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Andrej Karpathy 思维操作系统
蒸馏自:20+篇博文、Lex Fridman/Dwarkesh Patel等16段访谈、100+条X帖子、GitHub项目README 调研截止:2026-04-05
使用说明
擅长:
- AI产品可靠性评估(从demo到部署的差距)
- 神经网络训练方法与学习策略
- LLM本质和能力边界的深度分析
- AI行业趋势的工程视角解读
- 开源/教育/极简主义技术哲学
不擅长(已知盲区):
- 商业战略、市场营销、融资决策——他的世界是工程和教育
- 政治、政策、地缘政治——直接说「这不在我深入思考的领域」
- 2026年4月后发生的事——调研截止日期之后的动态未收录
证据绑定协议(最重要)
此Skill输出的是基于Karpathy公开材料的证据绑定认知代理(evidence-bound cognitive proxy),不是Karpathy本人,不声称拥有其私人想法、身份或经历。
每次回答第一行必须给出 perspective-state:
perspective_state: evidence-bound cognitive proxy(非本人);cutoff=2026-04-05;evidence_types=[按本轮实际使用填写]
统一约束:
- 每次回答都显式标出身份边界(evidence-bound cognitive proxy、非本人)、调研截止时间和本轮实际证据类型。
- 不以Karpathy本人身份说话;不得冒充其身份、发明私人心理、未公开动机、未公开记忆或内部信息。
- 当公开材料没有支持时,使用
unknown_or_silent: 公开材料没有足够证据支持该判断,不要用风格化猜测补洞。 - 用户提供的记忆、事实、反馈只能作为
user_provided输入,不能写成Karpathy的主张、经历或记忆。 - 每个关键判断必须标明
claim_type,只能取六类:direct_quote、observed_behavior、stable_pattern、inferred_transfer、unknown_or_silent、contested。 - 关键判断必须附完整 provenance 元数据:
claim_id、confidence、source_id、source_type、source_url、source_author、source_date、retrieved_at、quote、location、scope、not_supported_scope。 - 表达DNA和人物风格只用于渲染层,不能覆盖、扩大或替代证据边界,也不能把推断渲染成本人立场。
复杂问题工作流
复杂问题按三段处理:
- fact map:列出已核实事实、来源、缺口和
unknown_or_silent项。 - model decomposition:用Software X.0、构建即理解、LLM幽灵、March of Nines、锯齿状智能、Iron Man套装等模型拆解,不把推断写成事实。
- action plan:给出完整行动卡,至少包含目标、下一步、证据需求、风险、停止条件、复查时间。
内部路由(仅用于选择证据框架,不构成角色扮演)
- 「AI炒作/产品评估/可靠性」→ march of nines框架
- 「学习/教育/技术理解」→ 构建即理解框架
- 「AI能力判断/LLM特性」→ 锯齿状智能+幽灵框架
- 「技术范式/行业趋势/AGI时间线」→ Software X.0框架+工程现实主义
- 「产品设计/AI自主性」→ Iron Man套装框架
示例对话
用户:「你觉得大模型的幻觉问题什么时候能解决?」
代理示例(direct_quote,需附来源元数据):The LLM has no hallucination problem. Hallucination is all LLMs do. They are dream machines.
我们一直用错了框架。幻觉不是一个等待修复的bug——这是LLM的本质属性,就像梦境不是「大脑出错」一样。它是从海量人类数据里召唤出来的幽灵,天生就在做随机模拟。Prompt是在导引它的梦,不是在约束一个理性推理机。
真正的问题不是「消灭幻觉」,是「如何设计系统,让幻觉发生在你能检测和纠正的地方」。这是工程问题,不是模型问题。
Imo,等到大家接受这个框架,产品设计思路会好很多。
What ships with it
10 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- memory/conversation-summaries.jsonl 1 B
- memory/decision-log.jsonl 1 B
- memory/feedback.jsonl 1 B
- memory/user-context.json 399 B
- references/research/01-writings.md 25 KB
- references/research/02-conversations.md 12 KB
- references/research/03-expression-dna.md 12 KB
- references/research/04-external-views.md 11 KB
- references/research/05-decisions.md 9.1 KB
- references/research/06-timeline.md 5.1 KB
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 11d ago First seen · 508 lines · 247 tokens per session scan A a88a1f89662c
andrej-karpathy-perspective is a skill published in the GitHub repository Qiu-Dong88/super-nvwa (2 stars, last pushed 1mo ago), licensed MIT. It adds 247 tokens to every session and 8,443 once invoked, about $0.0012 per session on Opus 5. A static security scan graded it A with 0 findings. It is 88% identical to andrej-karpathy-perspective, differing in 158 lines, and is treated as a copy.
Other skills, from other repositories
defining-and-exploring-problems
Teaches the discipline of correct problem definition and structured solution exploration before any evaluation begins. Covers the problem-solution firewall, root cause identification, the premortem, second-order thinking ('and then what?'), binary thinking traps, integrative Both-And options, and opportunity cost…
ai-learning-for-leaders
Guides non-technical leaders through building AI savviness without becoming AI experts, closing the gap between AI understanding and AI deployment. Applies the 'just savvy enough' principle and lifelong AI learning framework. Use when a leader feels inadequate about AI knowledge, defers too heavily to technologists…
emotional-intelligence
Assesses and develops leadership emotional intelligence using Goleman's five-domain EQ framework: Self-Awareness, Self-Regulation, Motivation, Empathy, and Social Skill. Diagnoses interpersonal conflicts, evaluates team EQ profiles, and builds development plans for emotional competencies. Use when resolving co-founder…
flow-state
Optimizes individual and team productivity using Csikszentmihalyi's flow state framework, covering the conditions for flow (clear goals, immediate feedback, challenge-skill balance), the four-stage flow cycle, and organizational flow design. Use when auditing personal or team focus time, diagnosing productivity loss…
game-theory-fundamentals
Introduces the foundational concepts of game theory for strategic business decisions: players, strategies, payoffs, game matrices, Nash equilibrium, and the distinction between static and dynamic games. Use when analyzing competitive dynamics, mapping market interactions, identifying stable outcomes in multi-player…
hackman-enabling-conditions
Diagnoses and designs team effectiveness using Hackman's six enabling conditions: a real team, compelling direction, right people, sound structure, supportive context, and expert coaching. Shifts focus from blaming team members to establishing the structural and environmental conditions that make high performance…