prompt-analysis

提示词分析技能

使用本地 prompts.db SQLite 数据库分析 AI 提示词模式。

什么是 Git AI？

Git AI 是一款用于在 git 中追踪 AI 生成的代码和提示词的工具。它会存储：

• 每次 AI 对话（提示词和回复）
• 哪些代码行来自 AI，哪些经过人工编辑
• 接受率（保留了多少 AI 代码，以及修改了多少）
• 相关提交和作者

此技能会查询这些数据，帮助用户了解自己的 AI 编码模式。

初始化

首先，根据用户的问题确定范围：

用户提及	使用的标志
“我的提示词”或未指定	（默认值——当前用户、当前仓库）
“团队”“所有人”“所有作者”	--all-authors
某个具体人员的姓名	--author "<name>"
某个具体时间范围	--since <days>（默认值：30）

发现仅限于笔记——git-ai prompts 始终针对当前仓库运行（工作目录必须位于 git 仓库内）。若要分析多个仓库，请分别在每个仓库中运行命令。

运行初始化：

git-ai prompts [flags]

这会在当前目录中创建或更新 prompts.db。

架构参考

prompts 表包含：

• seq_id - 用于迭代的自增 ID
• id - 唯一提示词标识符
• tool - 使用的工具（例如，“claude-code”“cursor”）
• model - 模型名称（例如，“claude-sonnet-4-20250514”）
• human_author - 创建提示词的 Git 用户
• commit_sha - 相关提交（如果有）
• total_additions、total_deletions - 更改的代码行数
• accepted_lines、overridden_lines - 保留的行数与人工修改的行数
• accepted_rate - 比率：accepted / (accepted + overridden)
• messages - 对话的 JSON 数组
• start_time、last_time - Unix 时间戳

分析方法

针对汇总类问题（指标、比较）

使用直接 SQL 查询：

git-ai prompts exec "SELECT model, AVG(accepted_rate), COUNT(*) FROM prompts GROUP BY model"

针对单条提示词分析（分类、内容分析）

当问题需要检查每条提示词的内容（messages JSON）时，请使用子 Agent：

1. 在架构中添加分析列：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN work_type TEXT"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN analysis_notes TEXT"

2. 重置迭代指针：

git-ai prompts reset

3. 使用子 Agent 进行迭代——使用 subagent_type: "general-purpose"，通过 Task 工具启动并行子 Agent。每个子 Agent：

   • 运行 git-ai prompts next，获取一条提示词的 JSON
   • 分析 messages 内容
   • 更新数据库：git-ai prompts exec "UPDATE prompts SET work_type='...' WHERE id='...'"

4. 最终综合——查询已补充的数据：

git-ai prompts exec "SELECT work_type, COUNT(*), AVG(accepted_rate) FROM prompts GROUP BY work_type"

子 Agent 迭代模式

逐条处理提示词时，并行启动多个子 Agent。每个子 Agent 的提示词应包含：

Run `git-ai prompts next` to get the next prompt.

Analyze the messages JSON to determine: [specific analysis task]

Then update the database:
git-ai prompts exec "UPDATE prompts SET [column]='[value]' WHERE id='[prompt_id]'"

Return your analysis result.

每次启动 3-5 个子 Agent，检查结果，然后继续启动，直到 git-ai prompts next 返回“No more prompts”。

重要：git-ai prompts next 命令会以 JSON 形式返回分析所需的全部数据，包括：

• 包含完整对话（人工提示词和 AI 回复）的完整 messages 数组
• model、tool、accepted_rate、accepted_lines 等元数据

子 Agent 不应运行 git show 或 git log 等其他命令——分析所需的一切都包含在 git-ai prompts next 的 JSON 输出中。请明确告知子 Agent：

IMPORTANT: All data you need is in the JSON output from `git-ai prompts next`.
Do NOT run git commands. Analyze the `messages` field in the JSON directly.

迭代示例

示例 1：按工作类型对提示词分类

用户提问：“按工作类型（错误修复、功能、重构、文档）对我的提示词分类”

设置：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN work_type TEXT"
git-ai prompts reset

子 Agent 提示词：

Run `git-ai prompts next` to get the next prompt as JSON.

IMPORTANT: All data you need is in this JSON output. Do NOT run git commands.
Analyze the `messages` field directly.

Read the messages JSON and categorize this prompt into ONE of:
- "bug_fix" - fixing broken behavior, errors, or regressions
- "feature" - adding new functionality
- "refactor" - restructuring code without changing behavior
- "docs" - documentation, comments, READMEs
- "test" - adding or modifying tests
- "config" - configuration, build, CI/CD changes
- "other" - doesn't fit above categories

Update the database:
git-ai prompts exec "UPDATE prompts SET work_type='<category>' WHERE id='<prompt_id>'"

Return: the prompt id, your categorization, and a one-sentence reason.

综合查询：

SELECT work_type, COUNT(*) as count,
       ROUND(AVG(accepted_rate), 3) as avg_acceptance,
       SUM(accepted_lines) as total_lines
FROM prompts
WHERE work_type IS NOT NULL
GROUP BY work_type
ORDER BY count DESC

示例 2：分析提示词接受率较低的原因

用户提问：“为什么我的一些提示词接受率较低？”

设置：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN low_acceptance_reason TEXT"
git-ai prompts exec "UPDATE pointers SET current_seq_id = (SELECT MIN(seq_id) - 1 FROM prompts WHERE accepted_rate < 0.5 AND accepted_rate IS NOT NULL)"

子 Agent 提示词：

Run `git-ai prompts next` to get the next prompt as JSON.

IMPORTANT: All data you need is in this JSON output. Do NOT run git commands.
Analyze the `messages` field directly.

This prompt had a low acceptance rate (human modified most of the AI's code).
Analyze the messages JSON and identify the likely reason:
- "vague_request" - the prompt was unclear or underspecified
- "wrong_approach" - AI took a fundamentally wrong approach
- "style_mismatch" - code worked but didn't match project conventions
- "partial_solution" - AI only solved part of the problem
- "overengineered" - AI added unnecessary complexity
- "context_missing" - AI lacked necessary context about the codebase
- "other" - explain briefly

Update the database:
git-ai prompts exec "UPDATE prompts SET low_acceptance_reason='<reason>' WHERE id='<prompt_id>'"

Return: prompt id, the reason, and specific evidence from the conversation.

示例 3：找出可转化为可复用模式的提示词

用户提问：“我的哪些提示词解决了我可能再次遇到的问题？”

设置：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN reusable_pattern TEXT"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN pattern_description TEXT"
git-ai prompts reset

子 Agent 提示词：

Run `git-ai prompts next` to get the next prompt as JSON.

IMPORTANT: All data you need is in this JSON output. Do NOT run git commands.
Analyze the `messages` field directly.

Analyze whether this prompt represents a reusable pattern worth saving:
- Look for: common coding tasks, useful abstractions, clever solutions
- Skip: one-off fixes, highly context-specific changes, trivial edits

If reusable, set reusable_pattern to a short name (e.g., "api_error_handling", "form_validation", "test_mocking")
and pattern_description to a one-sentence description of what it does.

If not reusable, set both to NULL.

git-ai prompts exec "UPDATE prompts SET reusable_pattern='<name>', pattern_description='<desc>' WHERE id='<prompt_id>'"

Return: prompt id and whether you marked it as reusable (with the pattern name if yes).

示例 4：评估提示词的质量和清晰度

用户提问：“我的提示词有多清晰？哪些提示词可以写得更好？”

设置：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN clarity_score INTEGER"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN clarity_feedback TEXT"
git-ai prompts reset

子 Agent 提示词：

Run `git-ai prompts next` to get the next prompt as JSON.

IMPORTANT: All data you need is in this JSON output. Do NOT run git commands.
Analyze the `messages` field directly.

Score the HUMAN's prompt clarity from 1-5:
5 = Crystal clear: specific goal, context provided, constraints stated
4 = Good: clear intent, minor ambiguities
3 = Adequate: understandable but missing helpful context
2 = Vague: required AI to make significant assumptions
1 = Unclear: AI had to guess what was wanted

Also provide brief feedback on how the prompt could be improved.

git-ai prompts exec "UPDATE prompts SET clarity_score=<1-5>, clarity_feedback='<feedback>' WHERE id='<prompt_id>'"

Return: prompt id, score, and your feedback.

综合查询：

SELECT clarity_score, COUNT(*) as count,
       ROUND(AVG(accepted_rate), 3) as avg_acceptance
FROM prompts
WHERE clarity_score IS NOT NULL
GROUP BY clarity_score
ORDER BY clarity_score DESC

示例 5：分析提示技巧与接受率的相关性

用户提问：“分析我的提示技巧与接受率之间的相关性”

设置：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN technique TEXT"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN technique_notes TEXT"
git-ai prompts reset

子 Agent 提示词：

Run `git-ai prompts next` to get the next prompt as JSON.

IMPORTANT: All data you need is in this JSON output. Do NOT run git commands.
Analyze the `messages` field directly.

Analyze the HUMAN's prompting technique in the messages. Identify which techniques were used:
- "example_driven" - provided examples of desired output or behavior
- "step_by_step" - broke down the request into steps or phases
- "context_heavy" - provided extensive background/context about the codebase
- "minimal" - terse, brief request with little context
- "iterative" - built up solution through back-and-forth refinement
- "constraint_focused" - emphasized what NOT to do or specific requirements
- "reference_based" - pointed to existing code/files to follow as patterns
- "multiple" - combined several techniques (list them in notes)

Also note any specific effective or ineffective patterns in technique_notes.

git-ai prompts exec "UPDATE prompts SET technique='<technique>', technique_notes='<notes>' WHERE id='<prompt_id>'"

Return: prompt id, technique identified, the acceptance_rate, and your observations.

综合查询：

SELECT technique,
       COUNT(*) as count,
       ROUND(AVG(accepted_rate), 3) as avg_acceptance,
       ROUND(MIN(accepted_rate), 3) as min_acceptance,
       ROUND(MAX(accepted_rate), 3) as max_acceptance
FROM prompts
WHERE technique IS NOT NULL AND accepted_rate IS NOT NULL
GROUP BY technique
ORDER BY avg_acceptance DESC

示例 6：对失败的提示词进行定性分析

用户提问：“对失败的提示词（接受率为 0）进行定性分析”

“失败的提示词”的定义：

• accepted_lines IS NULL - 进行了对话，但没有代码被提交
• accepted_lines = 0 - 生成了代码，但被人类全部拒绝或重写

设置：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN failure_analysis TEXT"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN failure_category TEXT"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN improvement_suggestion TEXT"
-- Set pointer to iterate only over failed prompts
git-ai prompts exec "DELETE FROM pointers WHERE name='default'"
git-ai prompts exec "INSERT INTO pointers (name, current_seq_id) VALUES ('default', (SELECT MIN(seq_id) - 1 FROM prompts WHERE accepted_lines IS NULL OR accepted_lines = 0))"

子 Agent 提示词：

Run `git-ai prompts next` to get the next prompt as JSON.

IMPORTANT: All data you need is in this JSON output. Do NOT run git commands.
Analyze the `messages` field directly.

This prompt resulted in 0% acceptance - the human rejected or completely rewrote all AI-generated code.

Perform a qualitative analysis by reading the full conversation in messages JSON:

1. **What was requested?** Summarize the human's goal.
2. **What did the AI produce?** Summarize what code/solution was generated.
3. **Why did it fail?** Categorize into:
   - "misunderstood_intent" - AI solved the wrong problem
   - "poor_code_quality" - bugs, errors, or broken code
   - "wrong_technology" - used wrong framework/library/approach
   - "incomplete" - only partially addressed the request
   - "style_violation" - worked but violated project conventions
   - "overcomplicated" - solution far more complex than needed
   - "security_issue" - introduced vulnerabilities
   - "abandoned" - user changed direction mid-conversation

4. **How could the prompt be improved?** Specific, actionable suggestion.

git-ai prompts exec "UPDATE prompts SET failure_category='<category>', failure_analysis='<analysis>', improvement_suggestion='<suggestion>' WHERE id='<prompt_id>'"

Return a detailed analysis including: prompt id, what was requested, what went wrong, and how to prompt better next time.

综合分析：

SELECT failure_category,
       COUNT(*) as count,
       GROUP_CONCAT(id, ', ') as prompt_ids
FROM prompts
WHERE failure_category IS NOT NULL
GROUP BY failure_category
ORDER BY count DESC

然后查看各个 failure_analysis 和 improvement_suggestion 的值，总结经验教训。

示例 7：根据最佳实践对提示词评分

用户提问：“根据最佳实践为我的提示词评分”

设置：

git-ai prompts exec "ALTER TABLE prompts ADD COLUMN grade TEXT"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN grade_breakdown TEXT"
git-ai prompts exec "ALTER TABLE prompts ADD COLUMN grade_feedback TEXT"
git-ai prompts reset

子 Agent 提示词：

Run `git-ai prompts next` to get the next prompt as JSON.

IMPORTANT: All data you need is in this JSON output. Do NOT run git commands.
Analyze the `messages` field directly.

Grade the HUMAN's prompt against these best practices for AI-assisted coding:

**Grading Criteria (each worth 0-2 points):**

1. **Specificity (0-2)**: Does the prompt clearly state what needs to be done?
   - 0: Vague or ambiguous ("fix this", "make it better")
   - 1: General direction but missing details ("add validation")
   - 2: Specific outcome described ("add email validation that checks format and shows inline error")

2. **Context (0-2)**: Does the prompt provide necessary background?
   - 0: No context, assumes AI knows everything
   - 1: Some context but missing key details
   - 2: Relevant files, constraints, or project patterns mentioned

3. **Scope (0-2)**: Is the request appropriately sized?
   - 0: Too broad ("rewrite the app") or impossibly vague
   - 1: Large but manageable, could be broken down
   - 2: Focused, single-responsibility task

4. **Constraints (0-2)**: Are requirements and boundaries specified?
   - 0: No constraints given when they'd be helpful
   - 1: Some constraints but missing important ones
   - 2: Clear about what to do AND what not to do

5. **Testability (0-2)**: Can success be verified?
   - 0: No way to know if the result is correct
   - 1: Implicit success criteria
   - 2: Explicit expected behavior or acceptance criteria

**Calculate total (0-10) and assign grade:**
- A (9-10): Excellent prompt, follows best practices
- B (7-8): Good prompt, minor improvements possible
- C (5-6): Adequate prompt, notable gaps
- D (3-4): Poor prompt, significant issues
- F (0-2): Ineffective prompt, needs major revision

Store breakdown as JSON: {"specificity": X, "context": X, "scope": X, "constraints": X, "testability": X}

git-ai prompts exec "UPDATE prompts SET grade='<A-F>', grade_breakdown='<json>', grade_feedback='<specific feedback>' WHERE id='<prompt_id>'"

Return: prompt id, grade, total score, and one key improvement suggestion.

综合查询：

SELECT grade,
       COUNT(*) as count,
       ROUND(AVG(accepted_rate), 3) as avg_acceptance,
       SUM(accepted_lines) as total_lines_accepted
FROM prompts
WHERE grade IS NOT NULL AND accepted_rate IS NOT NULL
GROUP BY grade
ORDER BY grade

相关性分析：

-- See if grade correlates with acceptance rate
SELECT
    grade,
    ROUND(AVG(accepted_rate), 3) as avg_acceptance,
    COUNT(*) as n
FROM prompts
WHERE grade IS NOT NULL AND accepted_rate IS NOT NULL
GROUP BY grade
ORDER BY
    CASE grade
        WHEN 'A' THEN 1
        WHEN 'B' THEN 2
        WHEN 'C' THEN 3
        WHEN 'D' THEN 4
        WHEN 'F' THEN 5
    END

查询示例

按模型统计接受率：

SELECT model,
       ROUND(AVG(accepted_rate), 3) as avg_rate,
       SUM(accepted_lines) as total_accepted,
       COUNT(*) as prompt_count
FROM prompts
GROUP BY model
ORDER BY avg_rate DESC

最具成效的提示词（接受率高且代码行数高）：

SELECT id, tool, model, accepted_lines, accepted_rate,
       (last_time - start_time) as duration_secs
FROM prompts
WHERE accepted_rate > 0.8 AND accepted_lines > 50
ORDER BY accepted_lines DESC
LIMIT 10

按一天中的时段统计提示词：

SELECT
  CASE
    WHEN (start_time % 86400) / 3600 < 12 THEN 'morning'
    WHEN (start_time % 86400) / 3600 < 17 THEN 'afternoon'
    ELSE 'evening'
  END as time_of_day,
  COUNT(*) as count,
  AVG(accepted_rate) as avg_rate
FROM prompts
GROUP BY time_of_day

清理

prompts.db 文件是临时的。随时删除即可重新开始：

rm prompts.db

故障排除

如果数据库中没有提示词： 此代码库中的 Git AI 可能未正确设置。请指导用户：

• 查看 https://usegitai.com/docs/cli 了解设置说明
• 确认已安装并配置 git-ai
• 确保设置完成后使用 AI 辅助进行过提交

重要：Git AI 只会追踪之后的数据。 Git AI 无法追溯归类或追踪安装之前的提示词。只有在设置 Git AI 后进行的提交才会包含提示词数据。如果用户刚安装 Git AI，在积累更多 AI 辅助提交之前，可用数据可能有限。

注意事项和数据说明

并非所有提示词都有统计数据或提交记录：

• 如果进行了提示词对话，但没有提交代码，commit_sha 为 NULL
• 对于未提交的工作，accepted_lines、overridden_lines、accepted_rate 可能为 NULL 或 0
• 出于同样的原因，total_additions、total_deletions 可能为 NULL

原因：

• 用户进行了对话，但没有接受生成的代码
• 用户接受了代码，但从未提交
• 用户仍在进行工作（代码尚未提交）

“失败的提示词”的定义： 分析失败或未成功的提示词时，请使用以下定义：

• accepted_lines IS NULL - 进行了对话，但没有代码被提交
• accepted_lines = 0 - 生成了代码，但被人类全部拒绝或重写

查询：WHERE accepted_lines IS NULL OR accepted_lines = 0

在查询中处理：

-- Exclude uncommitted prompts from acceptance rate analysis
SELECT model, AVG(accepted_rate)
FROM prompts
WHERE commit_sha IS NOT NULL AND accepted_rate IS NOT NULL
GROUP BY model

-- Find prompts with conversation but no committed code
SELECT id, tool, start_time
FROM prompts
WHERE commit_sha IS NULL

重要说明

• messages 列包含 JSON——解析后即可读取对话内容
• 任何需要读取消息内容的分析都应使用子 Agent
• SQLite 是临时工作区——可根据分析需要自由添加列
• 默认范围较保守（当前用户、当前代码库、30 天）
• 只有用户明确提到团队或所有代码库时，才扩大范围

权限设置

为避免每次运行 git-ai 命令时都收到提示，请将以下内容添加到项目或用户设置中：

项目： .claude/settings.json 用户： ~/.claude/settings.json

{
  "permissions": {
    "allow": [
      "Bash(git-ai:*)"
    ]
  }
}

这对于子 Agent 迭代尤其重要，因为子 Agent 不会继承技能级权限。