Diagnosing Your Prompting Habits from AI Conversation History with prompt-review
機械翻訳 / Machine-translated
cccskills
9時間前・22分で読める
Original article: Diagnosing Your Prompting Habits from AI Conversation History with prompt-review (cccskills)
Skill page: prompt-review
Conclusions from this article
With prompt-review installed, you can receive a Japanese-language report on your prompting tendencies, technical comprehension, and degree of AI dependency — drawn from conversation history across 10 AI tools including Claude Code and Cursor.
This article explains the contents of the report, how to install and invoke it, how to filter by period or project, and what to check before using it.
The typical way to invoke it in Claude Code is:
/prompt-review 30
The skill first runs its bundled collection script, which automatically searches your computer for log files from each supported AI tool.
It extracts only your (user) prompts from the last 30 days of discovered history, and an agent reads and analyzes the content.
When finished, a Japanese-language report is created in a file named reports/prompt-review-YYYY-MM-DD.md in your working folder, and the path is reported back to you.
The report opens with a data source summary covering the detected tools and message counts.
Following that is a per-project summary listing each project's message count, tools used, and a one-line description of the work.
The report then continues with analysis across the following dimensions:
| Dimension | What the report covers |
|---|---|
| Technical comprehension map | Classifies mentioned technologies into three levels: "proficient," "basic understanding," and "still learning / unclear" |
| Prompting patterns | Effective phrasing, areas for improvement, habits like short acknowledgments and switching between Japanese and English |
| AI dependency | Whether you decide on approaches yourself or delegate that to AI; how you rely on AI for debugging and status checks |
| Growth trajectory | How prompt specificity has changed over time, newly adopted technologies, and recurring issues |
| Cross-project trends | Strengths and weak areas, differences by project type, and how you use different tools |
| Secret warnings | Lists any API keys found in history (masked), with a recommendation to reissue them |
Every dimension includes short quotations from your actual prompts as evidence.
Inferences are clearly labeled with phrasing such as "it is inferred that…"
When there is insufficient data for a dimension, the report honestly states: "Analysis was not possible due to insufficient data."
prompt-review is a skill for reflecting on how you make requests to AI, using your own conversation history.
It lets you re-read conversations that normally just flow by, through three lenses: technical comprehension, prompting patterns, and AI dependency.
The collection script supports the following 10 tools:
If you use multiple tools, any that are detected are analyzed together.
Each piece of information is labeled with its source in parentheses — e.g., "(Claude Code)" — so you can also see how your usage differs between tools.
The storage locations and formats for each tool's logs are documented in the bundled references/data-sources.md.
Technical comprehension is classified based on the phrasing of your prompts.
Technologies where you give imperative instructions with specific implementation details are classified as "proficient."
Technologies where you just name a concept with "please do X" are "basic understanding," and technologies where you ask "what is X?" or say "it's not working" fall into "still learning / unclear."
For AI dependency, the skill distinguishes between cases where you decide on an approach and only delegate the implementation, versus cases where you hand the decision-making itself to AI.
How you rely on AI for debugging and the frequency of requests like "check the status" are also part of the analysis.
Claude Code frequently asks "shall I do X?" and conversation history tends to accumulate many short replies like "y," "go ahead," or "please do."
This skill excludes from its comprehension and style analysis any affirmative responses or simple expressions of thanks that are roughly 20 characters or fewer.
Short messages that still contain specific instructions — like "30px please" or "make it async" — are retained.
This means the analysis is based on substantive requests.
The collection script also detects API keys, tokens, passwords, and connection strings embedded in prompts.
If any are found, a warning section appears immediately after the data source summary, recommending that the keys be reissued or revoked.
Any values shown in the report are masked — they are never written out in plain text.
Running the following command in your terminal adds the skill.
During installation, you can choose which agent to install it into, such as Claude Code or Codex.
npx skills add https://github.com/tokoroten/prompt-review --skill prompt-review
If you'd like the agent to install it for you, send the following message as-is.
It asks the agent to inspect the contents first, so you can verify what the bundled scripts do before they are placed.
次の Agent Skill をインストールしてください。
https://github.com/tokoroten/prompt-review/tree/main/.claude/skills/prompt-review
まず SKILL.md と付属ファイルの中身を確認し、不審な処理（外部への送信、認証情報の読み取りなど）がないことを確かめてから、Claude Code なら ~/.claude/skills/prompt-review/、Codex なら ~/.agents/skills/prompt-review/ に配置してください。
The installation location depends on which agent you use and the desired scope:
| Agent | Use across all projects | Use in a specific project only |
|---|---|---|
| Claude Code | ~/.claude/skills/ | .claude/skills/ in the project |
| Codex | ~/.agents/skills/ | .agents/skills/ in the project |
The distribution repository is https://github.com/tokoroten/prompt-review.
You can append a number or project name when invoking it to change the scope of analysis.
With nothing appended, all projects from the past 7 days are analyzed.
| Invocation | Scope |
|---|---|
| /prompt-review | All projects, past 7 days |
| /prompt-review 30 | Past 30 days |
| /prompt-review all (or 0) | All time |
| /prompt-review project-name | Filtered to projects with a partial name match |
| /prompt-review project-name 30 | That project, past 30 days |
/prompt-review
The basic form for reviewing how you made requests over the past week.
A cross-project analysis is saved as a Markdown file in the reports/ folder.
The per-project summary also gives you an at-a-glance view of what work you used AI for that week.
/prompt-review all
This targets your entire history.
The longer time span gives the "growth trajectory" section more material to work with.
The advantage is that changes in prompt specificity, newly adopted technologies, and recurring issues are organized chronologically.
/prompt-review myapp 30
Analyzes only projects whose names contain "myapp," limited to the past 30 days.
This is useful when you want to examine your comprehension and delegation style within a single development project.
After reading the report, try rewriting the prompts highlighted under "patterns for improvement" in your next request.
The four things this skill treats as effective patterns are: specific constraints, step-by-step instructions, sufficient context, and clearly stated expected output format.
If a vague request like "make it look nice" is flagged, start by adding whichever of those four elements is missing.
Before using it, check both the environment it runs in and the nature of the data being handled.
The collection script runs via the python command, so Python must be available in your environment.
Also, the skill is configured so that the agent does not invoke it automatically — you must call it explicitly with /prompt-review.
Pay attention to the script path.
The example commands in the instructions use the path ~/.claude/skills/prompt-review/ under your home directory.
However, other parts of the instructions ask you to use .claude/skills/prompt-review/ under your working folder.
If the agent says it can't find the script, tell it the actual location where you placed the files.
This skill reads logs from multiple local tools and runs a script via Bash.
Before installing, read scripts/collect.py and references/data-sources.md to confirm which files are accessed and that nothing is sent externally.
Collected prompts are saved as JSON in /tmp.
After analysis, deleting them is a safe practice.
The report is instructed to keep quotations short, replace personal names in file paths with <user>, and obscure project-specific confidential information.
Even so, work-related content may be readable, so review the full report before sharing it with your team.
If a secret warning is displayed, note that the original log still contains the key in plain text.
The estimated results are material for reconsidering how you make requests — they are not a judgment of your abilities.
License information could not be confirmed from the available materials, so please check the repository before using this in a professional context.
The skill's process goes only as far as reading your local logs and writing them to JSON in /tmp and Markdown in reports/.
There are no steps that send data to an external service.
However, since analysis is performed by the agent, the extracted prompts are read by the AI model you are using.
OpenAI Codex history is included as an analysis target, and the skill itself can be placed in ~/.agents/skills/.
However, the script paths written in the instructions assume the Claude Code installation location.
When running with Codex, tell the agent the actual path.
This skill is configured so that the agent does not invoke it automatically.
Type /prompt-review to invoke it explicitly.
Specify the period or project name after that.
Start with the default past 7 days (no arguments) to get a feel for the report format.
When you want to see your growth trajectory, specify a longer period such as all or 90.
prompt-review is a skill that reviews how you make requests to AI — drawing from conversation history across 10 tools including Claude Code and Cursor — and delivers a Japanese-language report.
Your level of comprehension per technology, your prompting habits, and how much you delegate to AI are all presented with quotations from your actual prompts.
Use the results as material for rewriting your next request, and review the full content before sharing it with others.
For installation instructions and the latest information, visit the cccskills page.
This article is a reprint of the cccskills feature "Diagnosing Your Prompting Habits from AI Conversation History with prompt-review."
Details about the skill and installation instructions can be found on the cccskills prompt-review page.