把“这篇论文在干什么、创新点是什么、机制是否成立、有什么值得迁移”变成可调度、可查证、可压缩的研究流程。
explore-skill 是一个面向 Codex / Agent Skills 体系的研究分析 skill。它不是普通“总结论文”的提示词，而是一套让主 agent 管理多轮调研、子 agent 独立查证、最终压缩成研究笔记的工作流。
它适合处理：
| 场景 | 它会做什么 |
|---|---|
| 论文理解 | 先读懂文章主线、章节结构、核心 claim 和证据组织 |
| 创新点分析 | 拆成 answer units，逐项检查旧痛点、修改点、机制、证据、边界 |
| 机制解释 | 建立因果链、替代解释、失败条件和理论依据 |
| 系统设计分析 | 抽模块、数据流、设计取舍、可迁移原则 |
| 复杂组合问题 | 先拆用户问题，再判断每个回答焦点是否已经足够回答 |
| 深度调研 | 用三轮结构管理子 agent，而不是无限递归 |
先拆问题，再查证据，最后压缩成自己的知识结构。
The skill forces the agent to avoid the common failure mode:
Round 1 看起来材料很多，于是直接写最终答案。
Instead, it uses:
User Question Decomposition：先拆用户真正问了几个问题；answer units：把创新点、机制、claim、模块等变成可检查对象；evidence ledger：记录每个结论的证据来源、缺口和状态；three-round workflow：只保留 Round 1 / Round 2 / Round 3，避免流程膨胀；stop-challenge：最终前检查是否过早停止。
flowchart TD
A[User Question] --> B[User Question Decomposition]
B --> C[Focus Table: overview / standard / deep]
C --> D[Round 1: Orientation]
D --> E[Checkpoint: Answer Units + Evidence Ledger]
E --> F{Can every focus be answered?}
F -- No --> G[Round 2: Focused Investigation]
G --> E
F -- Yes --> H[Round 3: Final Check]
H --> I{Early-stop risk or unsupported claim?}
I -- High-impact gap --> G
I -- Weak but answerable --> J[Degrade / Mark Weak]
I -- Supported --> K[Final Research Note]
J --> K
| Round | Role | What Happens |
|---|---|---|
round1/ |
Orientation | Read source material, understand paper/topic/system, collect candidate answer units |
round2/ |
Focused Investigation | Only research promoted answer units; no broad background expansion |
round3/ |
Final Check | Claim checking, evidence sufficiency, anti-early-stop audit, final synthesis |
The rule is intentionally strict:
Round 2 cannot open new broad research directions. Round 3 cannot become another literature review.
The same paper can require different depth depending on the user question.
| User asks | Focus depth | Expected behavior |
|---|---|---|
| “这篇论文在干什么？” | overview |
Round 1 is often enough |
| “这篇论文在干什么？创新点是什么？” | standard |
Round 1 may be enough, but evidence strength must be recorded |
| “深入分析每个创新点，机制和证据如何？” | deep |
Each innovation becomes an answer unit and usually enters Round 2 |
| “这个 claim 是否真的成立？” | deep |
Requires evidence ledger and Round 3 claim checking |
| “有什么能迁移到我的系统？” | standard or deep |
Requires applicability, boundary, and risk analysis |
explore-skill/
├── README.md
├── SKILL.md
└── .gitignore
During actual use, the agent may create temporary research artifacts in the working project:
tmp/
research_state.md
evidence_mapping.md
evidence_ledger.md
followup_candidates.md
stop_challenge.md
01-<short-topic>问题研究/
research.md
rounds/
round1/
round2/
round3/
Those generated artifacts are intentionally not part of this skill repository.
Copy this folder into your Codex skills directory:
Copy-Item -Recurse -Force .\explore-skill "$env:USERPROFILE\.codex\skills\explore-skill"Restart Codex so the skill metadata is reloaded.
The only required file is:
SKILL.md
The frontmatter contains the trigger metadata, and the body contains the workflow instructions.
用 explore-skill 分析这篇论文在干什么，创新点是什么：
https://arxiv.org/pdf/2604.00969
深入分析这篇论文的每个创新点：它解决了什么痛点，机制为什么成立，证据是否充分，有什么可以迁移到我的系统设计里？
这个方法的核心机制是否真的被实验支撑？请区分原文直接证据、综合推断和不确定部分。
The final note is organized by the user's actual question focus, not by sub-agent order.
## Paper / Topic
本质：
- 一句话本质
## 这篇文章在干什么
...
## 核心创新点
### Innovation 1
旧痛点 -> 修改点 -> 机制 -> 证据 -> 边界
### Innovation 2
...
## 证据强度与不确定部分
...
## 可迁移启发
...| Principle | Meaning |
|---|---|
| User-question first | The agent must first split the user's original question into answer focuses |
| Evidence before confidence | “我觉得够了” is not a stopping condition |
| Flat orchestration | Sub agents cannot create sub-sub agents; only the main agent schedules rounds |
| Three rounds only | Gap filling and verification are actions inside Round 2 / Round 3, not extra stages |
| Weak is allowed | Unsupported or partial claims can remain, but must be marked as weak |
| Notes over dumps | The final output compresses research into reusable knowledge, not raw retrieval logs |
Large language models often stop early when the first round of evidence feels coherent. For research work, that is not enough. This skill makes the agent externalize:
- what the user actually asked;
- which answer units are required;
- what evidence supports each unit;
- where the evidence is weak;
- why final output is allowed.
The goal is not to make every task heavy. The goal is to make depth proportional to the question.
This README follows common patterns from public agent-skill repositories:
- clear skill definition and folder expectations, as seen in Anthropic's public Agent Skills repository;
- curated-skill README structure from Codex and Claude skill collections;
- explicit installation, examples, workflow, and quality gates for human readers.
Useful references:
- Anthropic Agent Skills: https://github.com/anthropics/skills
- Awesome Codex Skills: https://github.com/composiohq/awesome-codex-skills
- Codex Skills Library: https://github.com/proflead/codex-skills-library
This skill is actively evolving around real research workflows:
- paper innovation analysis;
- multi-question decomposition;
- anti-early-stop checks;
- evidence ledger design;
- note-style compression.
Use it, test it on difficult papers, and tighten the workflow where the agent still stops too early.