Hubbestaiskill.com

Home / All Skills / Workflow & Core Tools / paul-graham-perspective

Workflow & Core Tools · alchaincyf/nuwa-skill

paul-graham-perspective

Paul Graham's thinking framework and mode of expression. Based on deep research covering 200+ essays, 12 podcasts/interviews, Twitter/X analysis, perspectives from 7 key critics, and a complete life timeline, it distills 5 core mental models, 8 decision heuristics, and a comprehensive expressive DNA. Purpose: Acts as a thinking advisor, analyzing startups, writing, product, and life decisions through PG's lens. Trigger when the user mentions "from PG's perspective", "what would Paul Graham think", "PG mode", or "paul graham perspective". It should also trigger even if the user simply says "help me think from PG's angle", "what would PG do", or "switch to PG".

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

Roleplay shell: Directly responds in the first person as Paul Graham once activated, stating a disclaimer only on first activation, and drops character immediately upon matching an exit keyword.

examples/paul-graham-perspective/SKILL.md

**此Skill激活后，直接以Paul Graham的身份回应。**

What it does: Turns the "PG perspective" into a toggleable immersive persona; uses a one-time disclaimer + EXIT TRIGGER constraints to maintain immersion while providing an exit channel, sharing the framework architecture with the nuwa series in the same repository.

Enforced Agentic protocol: First identifies question type; questions involving specific companies, individuals, or post-2024 events must undergo online research before answering via mental models.

examples/paul-graham-perspective/SKILL.md

**核心原则：PG不凭感觉说话。他写essay之前会做大量研究和思考。这个Skill也必须这样。**

What it does: Encodes "research before writing" as an unskippable Step 1 Classification → Step 2 WebSearch → Step 3 Answer; research summaries are organized internally and not output to the user.

CHECKPOINT three questions: Forces self-checks between steps on classification accuracy and whether it is hallucinating knowledge from training corpora, mandating headlines instead of fluff.

examples/paul-graham-perspective/SKILL.md

1. 第一句话是判断还是铺垫？如果是铺垫 → 砍掉，第一句必须是 headline。

What it does: Converts PG's writing discipline (conclusion first, honest hesitation, open-ended conclusions) into runtime forced self-checks to prevent model degradation.

9-rule failure mode Fallback tree, detailing fallback actions and fallback scripts for each case (including "ask the user for 3 key facts if data is missing").

examples/paul-graham-perspective/SKILL.md

| 9 | 4 段输出没给一个明确判断（全是「on one hand...on the other hand」） | 砍铺垫，第一句必须是 headline 判断 | 先结论后铺垫，PG 不做两面论 |

What it does: Equips each failure mode with a fallback action, converting "search yields nothing" into requesting funding round, user scale, and founder background from the user instead of hallucinating.

6 counter-example blacklists, explicitly forbidding the most common imitation errors (third-person self-citation, academic jargon, five-paragraph essays, listicles, fake familiarity).

examples/paul-graham-perspective/SKILL.md

| 5 | 给「5 个 tips」「10 条建议」列表 | PG 输出是 essay 不是 listicle，他自己原文骂过「listicles are cheeseburgers」 | 用 1-2 个核心判断 + 类比展开 |

What it does: Hardcodes three columns of "bad example / why it's wrong / correct approach", turning PG's writing preferences into hard constraints.

Scenario-to-model lookup table + response structure: Prioritizes models and heuristics by problem type, providing a "Reflect/Refocus → Core Thesis → Example → Caveats → No Summary" framework.

examples/paul-graham-perspective/SKILL.md

**多模型冲突时**：以「对用户当前决策最有行动指导意义」的模型为主，其他作补充视角。

What it does: Front-loads "which model to use" as a lookup table action, avoiding random invocation or stacking multiple concepts.

5 core mental models: Writing=Thinking, Taste as Cognitive Instrument, Iterative Discovery, Superlinear Returns, and Independent Thinking as Survival.

examples/paul-graham-perspective/SKILL.md

**一句话**：写作不是把想好的东西记下来，写作本身就是思考过程。

What it does: Provides "one sentence + evidence (including essay excerpts) + application + limitations" for each model; e.g., Model 4 uses 1% vs. 5% weekly growth to demonstrate superlinear returns.

8 decision heuristics: Fund People Not Ideas, Make Something People Want, Do Things That Don't Scale, Default Alive or Default Dead, Stay Upwind, Keep Your Identity Small, Maker's Schedule, and Am I Surprising Myself.

examples/paul-graham-perspective/SKILL.md

1. **Fund People Not Ideas**：在早期阶段，创始人的品质比idea重要100倍。

What it does: Pairs each heuristic with Viaweb/YC/personal experience case studies, turning principles into reproducible judgment actions.

Expressive DNA chapter: Sentence structures, 4 opening patterns, high-frequency templates, vocabulary taboos, certainty spectrum, and humor styles reproduce PG's prose style.

examples/paul-graham-perspective/SKILL.md

- **词汇禁忌**：绝不用delve、burgeoning、utilize、facilitate、methodology。绝不用学术黑话。绝不堆形容词。

What it does: Encodes "how to speak" into rules accompanied by original PG quotes and a certainty spectrum (decisive on facts, cautious on inferences) to ensure reproducible style.

5 honesty boundaries: Chicken sexer problem, Silicon Valley-centric bias, 2005-2014 experience obsolescence, public expression ≠ true thoughts, and research cutoff 2026-04.

examples/paul-graham-perspective/SKILL.md

2. **Silicon Valley中心视角**：我的框架建立在硅谷创业生态上。

What it does: Clearly distinguishes between "simulatable analytical frameworks" and "unreplicable intuitive judgment", acknowledging blind spots exposed by the Delve incident.

2Core Capabilities

01Writing = Thinking: writing is thinking; inability to write clearly implies lack of clear thought

02Taste as a cognitive instrument: trainable intuition for making better judgments under incomplete information

03Iterative Discovery: great things are made, not designed

04Superlinear Returns: finding domains where doubling input yields quadruple or greater output

05Independent Thinking as Survival: guarding against identity labels and conformity, recognizing traps where "everyone agrees it's good"

06Startup decision heuristics library (Fund People Not Ideas, Make Something People Want, Do Things That Don't Scale, Default Alive)

07PG-style output constraints: certainty spectrum (decisive on facts, cautious on inferences), vocabulary taboos, essay structure

08Output fallbacks: 9 if-then failure modes and Fallback phrasing

3External Dependencies

Kind	Dependency / Endpoint
api	宿主 WebSearch / 联网研究（正文强制要求）

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: Assessment is based solely on the examples/paul-graham-perspective/ directory: SKILL.md + 6 research md files + FIDELITY.md are entirely plain-text instructions with no scripts, no credential/cookie/env access (the single hit for "credential" in §3 refers to Jessica being called YC's "secret weapon" in 02-conversations.md, not a credential read), no runtime network calls, and no disk writes. The 195 URLs scanned in §3 are located in the "source/URL" annotations of research documents (paulgraham.com, wikipedia, medium, etc.), serving as cited source references rather than runtime endpoints, exempt from classification under Batch 4 guidelines. SKILL.md instructing the agent to use host WebSearch is a host-level action and not subject to grading.

5Second-Pass Verification

[ok] 5 心智模型 / 8 决策启发式是否在正文可落地 — 每个模型给「一句话 + 证据（含 essay 原文）+ 应用 + 局限」，8 条启发式各配 Viaweb/YC/个人经历案例。
[ok] 是否真的无脚本执行与写盘 — §3 的「執行外部程序」「寫盤 / 產生檔案」「網路請求」三类均记零命中；§1 清单只有 8 个 .md 文件，无 .py/.sh。
[ok] §3 的 1 条「憑證」命中是否代表读取凭证 — 命中位于 02-conversations.md，为 Jessica 被称 YC「secret weapon」一词，属普通英文词，非凭证读取；故 credential_reads 为空。
[ok] 195 条 URL 命中是否会升档 — 全部位于 references/research/ 的来源/URL 标注，属被引用资料而非执行期端点，按批4 口径不判级。
[ok] 负面信息是否被隐瞒或美化 — 诚实边界 5 条主动写入鸡性别鉴定师问题、硅谷中心视角、2005-2014 经验可能过时与「几乎从不说 I was wrong」，并列出 4 对内在矛盾（含 Delve 事件盲点），未做单向美化。
[ok] meta.installs 是否可定位单一 skill — 无 per-skill 安装量可定位，installs 保持 null。

6Verdict & Scenarios

Centers on "Writing is Thinking" as the core methodology and embeds warnings against writing degradation in the AI era into the model, offering sharper thematic focus than peer skills

Provides an additional "scenario-to-model lookup table" and "response structure" skeleton, yielding better output organization than other skills in the series

Style constraints drill down to specific wordlists (forbidding delve, burgeoning, etc.), ensuring high reproducibility

Honesty boundaries acknowledge that core judgment cannot be reproduced and explicitly document blind spots from the Delve incident

Best for: Best for: ① Structured thinking on early-stage startup directions and founder evaluation; ② Writing/expression practice (refining thoughts until they can be articulated clearly); ③ Product and growth judgments (Make Something People Want, Do Things That Don't Scale); ④ Career and life decisions (Stay Upwind, Keep Identity Small).

Not for: Not suitable for: ① Direct application by founders in non-tech/non-English markets or non-elite backgrounds (Silicon Valley-centric bias); ② Those seeking concrete investment target advice (this skill only provides frameworks); ③ Conversations requiring emotional empathy or support; ④ Contexts where first-person roleplay is unacceptable.

Install Agent-Ready

① Verified Mirror 2026-10-10

Option A · Direct DownloadDownload paul-graham-perspective.tar.gz

sha256: ad76c617160f2a77…

Option B · Copy Guide for Agent

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit fe03746870.