AI-Interview Skill 系统：Context-Aware 方法论加载机制#

1
prompt = base_prompt + 大量_skill_文档  # 一次性全部注入

1
if ContextAwareSkillLoader 检测到需要 skill:
2
    prompt = base_prompt + 匹配的 skill（按需加载）
3
else:
4
    prompt = base_prompt  # 无额外开销

1
class ContextAwareSkillLoader:
2

3
    def get_skills_for_context(self, agent, phase, action, state):
4
        matched_skills = []
5

6
        for skill_dir in [common/, agent/]:
7
            for skill_file in skill_dir.glob("**/SKILL.md"):
8
                skill = self._parse_skill(skill_file)
9

10
                # 关键：检测逻辑
11
                if self._matches_triggers(skill, phase, action, state):
12
                    matched_skills.append(skill)
13

14
        return matched_skills

1
def _matches_triggers(self, skill, phase, action, state):
2
    for trigger in skill.triggers:
3

4
        # 1. Phase 触发 - 当前阶段是否需要此 skill
5
        if "phase" in trigger and trigger["phase"] == phase:
6
            return True
7

8
        # 2. Action 触发 - 当前动作是否需要此 skill
9
        if "action" in trigger and trigger["action"] == action:
10
            return True
11

12
        # 3. Condition 触发 - 状态条件是否满足
13
        if "condition" in trigger and self._eval_condition(trigger["condition"], state):
14
            return True
15

16
    return False

1
---
2
name: 追问生成策略
3
description: 根据偏差分数决定是否追问
4
version: 1.0.0
5
agent: question
6
triggers:
7
  - action: generate_followup
8
  - phase: followup
9
---
10

11
# 追问生成策略
12

13
## 追问决策
14

15
当 deviation_score >= 0.6 且 followup_depth < max_followup_depth 时生成追问
16

17
## 追问方向
18

19
- dev < 0.3: 深入技术细节，说明具体实践
20
- 0.3 <= dev < 0.6: 引导扩展，补充项目背景
21
- dev >= 0.6: 肯定方向，探索相关经验

1
src/agent/skills/
2
├── common/                          # 通用技能（所有 Agent 共享）
3
│   ├── retry/                       # 重试策略
4
│   ├── error-handling/             # 错误处理
5
│   ├── llm-calling/                 # LLM 调用规范
6
│   └── state-management/            # 状态管理
7
│
8
├── orchestrator/                    # 编排协调
9
├── resume/                          # 简历解析
10
├── question/                        # 问题生成
11
│   ├── generation/                  # 问题生成策略
12
│   ├── followup/                    # 追问生成
13
│   └── deduplication/               # 去重检查
14
├── evaluate/                        # 评估
15
├── feedback/                         # 反馈
16
│   ├── correction/                  # 纠错反馈
17
│   ├── guidance/                    # 引导反馈
18
│   └── comment/                    # 点评反馈
19
├── review/                          # 审查
20
└── knowledge/                       # 知识管理

1
from dataclasses import asdict
2
from src.agent.skill_loader import skill_aware_prompt
3

4
async def generate_followup(state: InterviewState, ...):
5
    prompt = f"问题: {question}\n回答: {answer}\n"
6

7
    enhanced = skill_aware_prompt(
8
        agent="question",
9
        phase=state.phase,          # warmup / initial / followup
10
        action="generate_followup",
11
        base_prompt=prompt,
12
        state=asdict(state),        # 用于 condition 匹配
13
    )
14

15
    result = await invoke_llm(enhanced)

1
原始 prompt: 请基于以下问题生成追问...
2

3
## 参考方法论（按需使用）
4

5
### [question] 追问生成策略
6
# 追问生成策略
7

8
## 追问决策
9
deviation_score >= 0.8 → 继续深入
10
...
11

12
### [common] 重试策略
13
# 重试策略
14
...