Home / All Skills / Workflow & Core Tools / andrej-karpathy-perspective
English (EN)简体中文 (ZH)日本語 (JA)
Workflow & Core Tools · alchaincyf/nuwa-skill

andrej-karpathy-perspective

Andrej Karpathy's mental framework and communication style. Systematically distilled from 20+ blog posts, 16 in-depth interviews, and 100+ X posts, extracting 6 core mental models, 8 decision-making heuristics, comprehensive Chinese output adaptation, and a quick reference for signature sentence patterns. Usage: Serves as a thinking consultant, analyzing AI technical reliability, learning methodologies, industry trends, and product design from Karpathy's perspective. Triggered when users mention "from Karpathy's perspective," "how would Karpathy see this," "Kapathy," or "karpathy mode." Also applicable to discussions on Software 2.0/3.0, vibe coding, neural network training, AI hype assessment, and LLM capability boundaries. Can also be triggered if the user simply says "from an engineering realism perspective," "march of nines," "building is understanding," or "jagged intelligence." Does not trigger for generic AI-related questions—only activates when a Karpathy-style thinking framework is explicitly requested.

L1 Basic · Source registration onlyV1 Commit-PinnedDual-pass Audit
Author alchaincyfGitHub alchaincyf/nuwa-skill ↗License MIT（仓库根 LICENSE）Commit fe03746870
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

The entire skill is pure prompt-based roleplay: the directory contains only one SKILL.md and 6 research attachments, with no scripts or data pipelines; once activated, the model answers directly in Karpathy's first-person voice.

examples/andrej-karpathy-perspective/SKILL.md
**此Skill激活后，直接以Karpathy的身份回应。**
What is happening here: Capabilities stem entirely from prompt orchestration—triggered via frontmatter, constraining persona and output form via body text, without any executable code.

Anti-drift design: Outputs a disclaimer only once upon initial activation, and provides an explicit EXIT TRIGGER; the user can revert to normal mode simply by saying "exit / switch back to normal."

examples/andrej-karpathy-perspective/SKILL.md
🛑 **STOP（仅一次）**：首次激活时输出免责声明一次——「我以Karpathy视角和你聊，基于公开言论推断，非本人观点」。后续对话**绝不**重复。
What is happening here: A one-time disclaimer avoids repetition in every turn; the EXIT TRIGGER allows the user to terminate the roleplay at any time without getting stuck in character.

Triggering is determined entirely by Chinese/English trigger words in the description, explicitly excluding scenarios where the user is "just asking generic AI questions."

examples/andrej-karpathy-perspective/SKILL.md
当用户提到「用Karpathy的视角」「Karpathy会怎么看」「卡帕西」「karpathy模式」时使用。
What is happening here: Separates "explicitly wanting this persona framework" from "asking broad AI questions" to reduce false triggers.

The core is a three-step Agentic Protocol: classify the question first, enforce web research, and finally answer using the framework, with two CHECKPOINT gates established.

examples/andrej-karpathy-perspective/SKILL.md
| **需要事实的问题** | 涉及具体模型/产品/公司/技术细节/最新发布 | → 先研究再回答（Step 2） |
What is happening here: Addresses outdated training data: factual/hybrid questions must gather evidence first, while pure framework questions can be answered directly.

The research step is formulated as a non-skippable hard rule explicitly demanding that "tools (WebSearch, etc.) must be used"—the skill itself does not implement networking; execution relies on host tools.

examples/andrej-karpathy-perspective/SKILL.md
**⚠️ 必须使用工具（WebSearch等）获取真实信息，不可跳过。**
What is happening here: This is also where its external dependency lies: authenticity and timeliness rely on the host's search capabilities; the skill only dictates when search is mandatory and against which checklist.

Judgments are structured using six mental models: Software X.0, building is understanding, the summoned ghost, march of nines, jagged intelligence, and the Iron Man suit; subsequent decision heuristics and communication DNA are tied to these models.

examples/andrej-karpathy-perspective/SKILL.md
**一句话**：编程语言在历史上只发生过两次根本性变化，我们正处于第三次。
What is happening here: Each model follows a five-part structure of "one-liner / core argument / what he said / how to apply / limitations"; limitations are just as important as the arguments.

Eight decision heuristics plus an entire section on communication DNA (imo tags, short sentence pauses, banned corporate buzzwords), accompanied by a "Chinese output adaptation" mapping stylistic markers into functionally equivalent Chinese expressions.

examples/andrej-karpathy-perspective/SKILL.md
1. **时间轴拉长批评**：不直接否定「X年就能实现」的说法，而是把时间轴拉长——「这是这个十年的事，不是这一年的」
What is happening here: The Chinese adaptation table maps "imo" to "我觉得/说实话", "hands down" to "就是这个，没别的", etc., avoiding awkward literal translations.

Concludes with a 9-item Fallback tree and an 8-item anti-pattern blacklist, constraining research failures, identity provocation, hedging, quoting cliches, and similar cases.

examples/andrej-karpathy-perspective/SKILL.md
异常先识别再处理；绝不静默跳过、绝不装作了解没了解过的事、绝不在身份争辩里耗时间。
What is happening here: This explicitly engineers "how roleplay can fail": every anomaly has a "first-line fix + fallback" rather than leaving it to the model's improvisation.

Self-discloses honesty boundaries: timeliness, gaps between public statements vs. private thoughts, inability to replace Karpathy's own creativity, and explicitly states the research cutoff date.

examples/andrej-karpathy-perspective/SKILL.md
1. **时效性**：Karpathy的技术立场更新极快（他2025年10月还说Agent无用，12月就转为80%使用）。本Skill基于2026年4月的信息，此后的动态未被捕捉。
What is happening here: Clarifies upfront that conclusions are valid only up to the research cutoff date, being transparent that real figures can change their stances.

2Core Capabilities

01A persona mode answering in Karpathy's first-person voice: activated via trigger words, one-time disclaimer, supports explicit exit; uses "I" instead of "Karpathy would think."
02An agentic "research before answering" pipeline: three question categories + research checklists for two question types + two CHECKPOINT gates.
03Six mental models (Software X.0 / building is understanding / the summoned ghost / march of nines / jagged intelligence / Iron Man suit) serving as the analytical framework.
04Eight actionable decision heuristics (zoom out the timeline, build from scratch to verify, prioritize data flywheels, imo tags, don't be a hero, etc.).
05Communication DNA: sentence structure preferences, vocabulary traits, rhythm, expressions of certainty, humor style, including a Chinese output adaptation table.
06Anti-pattern blacklist and failure fallback tree to prevent impersonating stances, quote-stuffing, AI customer service tone, and forcing answers on out-of-scope questions.
07Personal timeline, intellectual lineage, core values, and explicit antipathies, used to align judgment tendencies.
08Honesty boundaries and research source tiering (primary/secondary), explicitly noting corpus cutoff as of 2026-04.

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Can be auto-loaded
For Administrators
Register source and version
Evidence & Triggers (File Paths & Excerpts)
  • SKILL.md: Pure prompt instructions and rule text; no script execution or external impact.
Audit Notes: Evaluation is based solely on the skill directory itself: consists entirely of pure Markdown instructions (SKILL.md + 6 reference files + FIDELITY.md). Risk scan shows 0 hits across "external program execution / disk writes / network requests / dangerous flags / sandbox escape". URL hits are merely external link references within the documentation, with 0 hits for credentials or environment variables. It reads no credentials, writes no files, and initiates no network requests, hence rated 🟢 green (pure instructions). Operational considerations (do not affect tier): web research is executed by host tools whose retrieval results may not be verifiable; first-person persona outputs may make unverified claims appear authoritative; research cutoff is 2026-04-05, and subsequent shifts in the figure's stances are not covered.

5Second-Pass Verification

  • [ok] 全部 evidence.quote 是否逐字存在于 bundle（回源复核） — 本 JSON 的每条 quote 均取自 bundle 内已展开文件的可见原文（SKILL.md 全文、06-timeline.md 等），未使用被截断段落中的不可见文字；quote 长度均远超 12 字符下限。
  • [ok] 目录清单与 internal_assets 是否一致 — §1 列举 8 个文件（FIDELITY.md、SKILL.md、6 份 references/research/*.md），internal_assets 逐条对应，无多列或少列。
  • [ok] frontmatter description 与 meta.official_desc 一致性 — 骨架 meta.official_desc 即 SKILL.md frontmatter description 的逐字内容（含 20+ 篇博文/16 段访谈/100+ 条帖子与全部触发短语），保留未改。
  • [ok] 是否存在未披露的脚本/网络/凭证行为 — 风险扫描显示「执行外部程序／写盘／网络请求／危险旗标／沙箱」均 0 命中，凭证 0 命中；确认 skill 目录内无可执行代码，WebSearch 仅出现在提示词层。
  • [ok] commit / 授权 / 调研时间等元数据 — skill.commit 与 audit.skill_commit 同为 fe0374687037c4cc51a65c1e0c145afe2981dc69；授权为仓库根 MIT；调研截止 2026-04-05，与 SKILL.md「调研截止：2026-04-05」相符。
  • [unlocatable] FIDELITY.md 正文内容 — §1 列出 examples/andrej-karpathy-perspective/FIDELITY.md（2781 bytes／20 行），但 bundle §2 未展开其正文，故本次不对其内容作任何声称，仅在 internal_assets 记录其存在与大小。
  • [unlocatable] references/research 的 01/02/03 部分内容 — bundle 对 02（截断 705 bytes）、03（截断 3262 bytes）等标注「已截断」，未展开部分本次未引用、未据以立论，避免把不可见文本当成证据。

6Verdict & Scenarios

  • Breaks down "a person's thinking" into an executable three-stage pipeline (question classification -> web evidence retrieval -> framework response), rather than a vague persona description.
  • Mental models and heuristics specify failure boundaries down to the "limitation" level (e.g., march of nines being overly strict for consumer-facing creative scenarios).
  • Engineered anti-drift: one-time disclaimer, explicit exit anchor, 9 fallback rules, and 8 anti-pattern blacklist items.
  • Unusually granular evidence tiering: direct quotes marked as "what he said", research files rated by source reliability, distinguishing "said by him / paraphrased / inferred".
  • Dedicated adaptation for Chinese usability instead of directly copying English catchphrases.
    • Best for: Best for: Users wanting to evaluate AI product reliability, learning paths, and AI trend forecasts through Karpathy's framework of engineering realism and learning methodologies; scenarios requiring structured "mental models + actionable heuristics" rather than motivational fluff; conversations tolerant of web retrieval latency. Also serves as an engineering exemplar of "how to translate a thinker's mindset into a persona skill complete with fallbacks and localized adaptation."
    Not for: Not suitable for: Users seeking Karpathy's authentic personal stance or latest real-time opinions (cutoff is 2026-04, and fundamentally inferred from public statements); environments requiring offline operation / lacking web search capabilities (core workflow mandates research); lightweight Q&A only seeking quick polishing or casual chat; commercial/political/policy-related questions (acknowledged by the skill as blind spots).
    Install Agent-Ready
    ① Verified Mirror 2026-10-10
    Option A · Direct DownloadDownload andrej-karpathy-perspective.tar.gz
    sha256: 95db3f22c1c5275b…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit fe03746870.