大语言模型

今天的路线图

LLM 到底在做什么？

因此，模型越像“懂了”，越需要问：它的判断来自上下文证据、训练中学到的模式，还是一次流畅的概率补全？

Token 与上下文窗口

Transformer：注意力机制

训练过程：从“会续写”到“会听指令”

为什么会幻觉？

从 LLM 到 Skill：工程化这一层

Prompt vs. Skill：差别在哪里？

Skills 机制总览

一个 Skill 最少需要什么？

skills/
  interview-coding/
    SKILL.md
    templates/
      codebook.md
    scripts/
      validate_labels.py
    examples/
      sample_interviews.csv

SKILL.md 的最小格式

---
name: interview-coding
description: Code interview excerpts into a finite qualitative codebook.
---

# Interview Coding Skill

Use this when the user asks to classify interview excerpts,
open-ended survey responses, or field notes.

Workflow:
1. Verify input path, encoding, and row count.
2. Read or create a finite codebook.
3. Label each excerpt with code, confidence, and evidence quote.
4. Save outputs as CSV and Markdown audit report.

为什么不是把全文都塞进系统提示词？

<available_skills>
  <skill>
    <name>interview-coding</name>
    <description>Code interview excerpts...</description>
    <location>/workspace/skills/interview-coding/SKILL.md</location>
  </skill>
</available_skills>

Skill 从哪里加载？优先级是什么？

Load-time filters：不是所有 skill 都会出现

---
name: policy-crawler
description: Crawl policy pages into auditable CSV.
metadata: {"openclaw":{"requires":{"bins":["python3"],"env":["TAVILY_API_KEY"]},"primaryEnv":"TAVILY_API_KEY"}}
---

openclaw.json：skills 的控制面

{
  skills: {
    allowBundled: ["gemini", "peekaboo"],
    load: {
      extraDirs: ["~/Projects/course-skills"],
      watch: true,
      watchDebounceMs: 250
    },
    install: {
      preferBrew: true,
      nodeManager: "npm"
    },
    entries: {
      "policy-crawler": {
        enabled: true,
        apiKey: { source: "env", provider: "default", id: "TAVILY_API_KEY" },
        config: { maxPages: 30 }
      },
      sag: { enabled: false }
    }
  }
}

Env / API Key 的作用域

Session snapshot 与 watcher

{ skills: { load: { watch: true, watchDebounceMs: 250 } } }

Plugin skills、ClawHub 与课堂边界

clawhub install <skill-slug>
clawhub update --all
clawhub sync --all

Skill 与 GPT 模型：不是同一层

GPT 系列：给 Agent 工作流的粗略选型

注：模型名与可用性会变化；正式部署前以当前 provider / OpenAI API 文档为准。

Reasoning effort：给 Skill 加一个“成本旋钮”

把模型策略写进 Skill，而不是写死模型

# Model policy

Default:
- Use a fast/low-effort model for schema checks, deduplication, and simple labeling.

Escalate:
- Use a stronger reasoning model when:
  1. codebook definitions conflict,
  2. confidence is low on more than 20% of rows,
  3. the task requires tool-heavy multi-step execution.

Do not:
- Change model silently after producing results.
- Hide cost/latency tradeoffs from the user.
- Treat model output as validated evidence without spot-checks.

Skill 设计协议：先把规格空间写有限

课堂示例：开放题编码 Skill

---
name: open-response-coding
description: Classify Chinese open-ended survey answers with a finite codebook and audit trail.
---

# Open Response Coding

Use when the user provides CSV/XLSX/text responses and asks for coding,
categorization, or qualitative labeling.

Rules:
- Verify file path, encoding, row count, and target column before analysis.
- Do not invent unlimited categories; propose at most 6 codes first.
- For each row, output: id, code, confidence, evidence_quote.
- Save `coded_responses.csv` and `coding_report.md`.
- Randomly spot-check 10 rows and report disagreements or low confidence.
- If input contains direct identifiers, stop and ask for de-identification.

把脚本和模板放进 Skill，而不是塞进 Prompt

Before final answer:
1. Run:
   python3 {baseDir}/scripts/validate_labels.py coded_responses.csv
2. Include validation output in the report.
3. If validation fails, repair the CSV instead of hiding the error.

测试一个 Skill：不要只看“有没有回复”

类型一：编码与标注 Skill

类型二：政策文本采集 Skill

Output contract:
policy_list.csv columns:
source, title, date, agency, url, fetched_at, hash

类型三：可复现分析 Skill

课堂练习：20 分钟写一个 Mini Skill

1. 1选择一个小任务：开放题分类、政策标题抽取、CSV 描述统计。
2. 2写 skills/my-mini-skill/SKILL.md：name、description、触发条件、输出契约。
3. 3加一个样例输入，要求 agent 保存结构化输出。
4. 4新开会话测试：确认它读了 skill，且没有绕开验证步骤。

常见失败模式

今天的 takeaway