prompt-review Skill

Analyzes the user's past AI Agent conversation histories, estimates technical understanding level, prompting patterns and AI dependency to generate a report. The report will be written in Japanese to

reports/prompt-review-YYYY-MM-DD.md

Argument Processing

Parse

$ARGUMENTS

and process parameters according to the following rules:

Numeric only → Days filter (example:
```
30
```
→ data of the past 30 days)
String only → Project name filter (partial match)
String + number → Project name + Days filter (example:
```
yonshogen 30
```
)
No arguments → Cross all projects, data of the past 7 days (default)

Step 1: Data Collection (Script Execution)

Run the preprocessing script scripts/collect.py to collect data. This script automatically detects logs from Claude Code, GitHub Copilot Chat, Cline, Roo Code, Windsurf, Antigravity, and returns filtered JSON to standard output.

Assemble script options from arguments

Parse

$ARGUMENTS

and run the following command in Bash:

bash

python ~/.claude/skills/prompt-review/scripts/collect.py [OPTIONS] > /tmp/prompt-review-data.json

No arguments → No options (default: past 7 days)
Numeric only (example:
```
30
```
) →
```
--days 30
```
```
all
```
or
```
0
```
→
```
--days 0
```
(all time period)
String only (example:
```
yonshogen
```
) →
```
--project yonshogen
```

String + number (example:

yonshogen 30

) →

--project yonshogen --days 30

Important: Do not use a relative path from this skill file for the script path. Use the absolute path

.claude/skills/prompt-review/scripts/collect.py

of the project where the skill is stored, specified relative to the current working directory (

cwd

Read Output

After the script runs, read

/tmp/prompt-review-data.json

with the Read tool.

Output JSON structure:

json

{
  "summary": {
    "total_messages": 2616,
    "detected_tools": ["Claude Code", "GitHub Copilot Chat"],
    "filter_days": null,
    "filter_project": null
  },
  "sources": [
    {
      "tool": "Claude Code",
      "status": "検出",
      "messages": [
        {"text": "プロンプト本文", "timestamp": "2025-09-29 03:16", "project": "yonshogen"}
      ],
      "period": "2025-09-29 03:16 〜 2026-03-12 04:58"
    }
  ],
  "project_stats": {
    "farbrain": {"count": 668, "tools": ["Claude Code"]},
    "yonshogen": {"count": 215, "tools": ["Claude Code"]}
  },
  "secret_warnings": [
    {
      "tool": "Claude Code",
      "project": "some-project",
      "timestamp": "2025-10-01 12:00",
      "type": "OpenAI API Key",
      "masked_value": "sk-abc12***xyz9",
      "prompt_excerpt": "APIキーはsk-abc123..."
    }
  ]
}

Step 2: Analysis

After reading

/tmp/prompt-review-data.json

with Read, analyze the user prompts in the

messages

array from the following perspectives. Be sure to include specific evidence (quotes of actual prompt fragments) for each perspective.

Preprocessing: Create project-specific summaries and exclude short responses

First, use

project_stats

to get an overview of message counts and used tools per project. Read the prompt content for each project, and summarize the work being done in that project in one line. This information will be output in the "2. Project-specific Summary" section of the report. Next, exclude short affirmative responses from analysis.

Preprocessing: Ignore short affirmative responses

Claude Code often asks users for confirmation like "Shall I do ~?", and short user affirmative responses to these are not valuable for analyzing prompting ability or technical proficiency. Exclude the following types of messages from analysis:

Simple affirmations: 「y」「yes」「はい」「うん」「ok」「sure」「yep」「yeah」
Execution instructions: 「進めて」「やって」「do it」「doit」「go」「go ahead」「proceed」
Approvals: 「それで」「それでいい」「それでお願いします」「お願いします」「いいよ」「いいです」「大丈夫」
Gratitude only: 「ありがとう」「ありがとうございます」「thanks」「thx」

Judgment criteria: Messages that are short (roughly 20 characters or less) and match the above patterns. However, even short messages that include specific technical instructions (e.g. "30pxがいいです", "asyncで") should not be excluded.

2a. Technical Proficiency Map

Extract technical concepts mentioned in prompts, and classify into 3 levels:

Proficient: Confident instructions, accurate terminology use, presentation of specific implementation policies
Basic understanding: Knows the concept but delegates details to AI
Learning/Ambiguous: Question format, misunderstandings present, lots of trial and error

Classification signals:

Specific instructions in imperative form → Proficient
"Please do ~" + concept name only → Basic understanding
"What is ~~?" "~~ is not working" "How should I do it?" → Learning

2b. Prompting Pattern Analysis

Effective patterns: Specific constraint specification, step-by-step instructions, sufficient context provision, clear specification of expected output format
Improvable patterns: Ambiguous instructions (e.g. "Make it nice"), insufficient context, confusion of purpose and means
Characteristic habits: Frequency of short approvals (「y」「doit」 etc.), use of Japanese/English

2c. AI Dependency Analysis

Per project/tool:

Cases where user decides policy and delegates implementation to AI (proactive)
Cases where user delegates even policy decision to AI (dependent)
Pattern of relying on AI for debugging/error resolution
Pattern of frequently requesting status checks from AI (e.g. "Check the status")

2d. Growth Trajectory

Time series analysis:

Changes in prompt quality/specificity
Newly adopted technical concepts
Recurring issue patterns

2e. Cross-project and cross-tool trends

Strong and weak areas
Differences in behavior by project type
How AI tools are used differently (if multiple tools are detected)

2f. Secret/Credential Warning

If the

secret_warnings

array is not empty, output a warning section at the beginning of the report (right after the data source summary). List all secrets such as API Keys, Tokens, Passwords, connection strings detected by the script, and recommend the user to rotate (reissue/invalidate) the keys.

Never write secret values in plain text in the report. Use the masked value (

masked_value

) returned by the script as-is.

Step 3: Report Generation

Generate the report in Japanese according to the template at references/report-template.md.

Output Rules

If the
```
reports/
```
directory does not exist, run
```
mkdir -p reports
```
in Bash
Use the Write tool to output to
```
reports/prompt-review-YYYY-MM-DD.md
```
(YYYYMMDD is the execution date)
After generating the report, notify the user of the file path

Writing Notes

The report must be output in Japanese without exception
Clearly state speculations as "It is estimated that ~" "There is a possibility of ~"
When quoting the original prompt, keep excerpts short and pay attention to privacy:
- Mask personal name parts of file paths to
```
<user>
```
- Redact project-specific confidential information
For parts that cannot be analyzed, honestly state "Could not analyze due to insufficient data"
Clearly indicate which tool the information comes from in parentheses (e.g. "(Claude Code)" "(Copilot Chat)")
Sections without evidence may be omitted

Reference Resources

scripts/collect.py — Data collection and preprocessing script (Python)
references/data-sources.md — Details of log storage locations and formats for each AI tool
references/report-template.md — Report structure template

prompt-review

NPX Install

Tags

SKILL.md Content (Chinese)