<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by driller on Medium]]></title>
        <description><![CDATA[Stories by driller on Medium]]></description>
        <link>https://medium.com/@patraqushe?source=rss-39e7a82c8009------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*1ph9jNIeiI8YdTawoBb1Ug.png</url>
            <title>Stories by driller on Medium</title>
            <link>https://medium.com/@patraqushe?source=rss-39e7a82c8009------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 09:21:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@patraqushe/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[hachimoku: A Multi-Agent Code Review CLI Defined in TOML]]></title>
            <link>https://medium.com/@patraqushe/hachimoku-a-multi-agent-code-review-cli-defined-in-toml-eee27dbf65f9?source=rss-39e7a82c8009------2</link>
            <guid isPermaLink="false">https://medium.com/p/eee27dbf65f9</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[code-review]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[github]]></category>
            <dc:creator><![CDATA[driller]]></dc:creator>
            <pubDate>Wed, 18 Feb 2026 13:29:15 GMT</pubDate>
            <atom:updated>2026-02-18T13:43:51.861Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TnIGtWJxz5w9LbqYBXjL_w.png" /></figure><p>Bring multiple expert eyes to every pull request with customizable review agents</p><h3>In the Age of Coding Agents, Code Review Can’t Keep Up</h3><p>Coding agents are churning out pull requests at an unprecedented pace, but the review process remains a human bottleneck. Bug detection, security analysis, test coverage evaluation, type safety checks — covering all of these perspectives as a single reviewer is a tall order. Reviews slow down the pipeline.</p><p>In a previous article, I introduced PR Review Toolkit, an official Claude Code plugin for automated code review. PR Review Toolkit is convenient for quick reviews within Claude Code, but it doesn’t support adding or customizing review perspectives.</p><p>That limitation led me to build <strong>hachimoku</strong>, a multi-agent code review CLI tool where agents are defined and customized through TOML files. It runs on Claude Code and works within the flat-rate plan.</p><p>GitHub: <a href="https://github.com/drillan/hachimoku">https://github.com/drillan/hachimoku</a></p><h3>What Is hachimoku?</h3><p>hachimoku (8moku) is a CLI tool that performs code reviews using multiple specialized agents. The name comes from the Japanese proverb “Okame Hachimoku” (岡目八目), meaning “a bystander with eight eyes sees more than the players.” The concept is that multiple eyes (agents) examine code from different angles.</p><p>Agent definitions are managed in TOML files, so you can add or customize review perspectives without changing any code.</p><p>Key features:</p><ul><li><strong>Three review modes</strong>: diff (branch diff), PR (GitHub PR), and file (specified files)</li><li><strong>Six built-in agents</strong> covering common review concerns</li><li><strong>Selector</strong>: Analyzes changes and automatically picks the right agents</li><li><strong>Aggregator</strong>: Deduplicates and merges findings from multiple agents, then generates prioritized action items</li><li><strong>TOML-based custom agents</strong>: Add project-specific review perspectives with a simple file</li><li><strong>Parallel or sequential execution</strong>: Choose the execution strategy that fits your workflow</li><li><strong>Markdown and JSON output</strong></li><li><strong>JSONL review history</strong>: Automatically accumulated for trend analysis</li></ul><h3>Built-in Agents</h3><p>hachimoku ships with six built-in agents, comparable to those in PR Review Toolkit:</p><p><strong>main phase:</strong></p><p>- <strong>code-reviewer: </strong>Overall review for code quality, bugs, and best practices</p><p>- <strong>silent-failure-hunter: </strong>Detection of silent failures and error handling issues</p><p>- <strong>pr-test-analyzer: </strong>Test coverage evaluation</p><p>- <strong>type-design-analyzer: </strong>Type design quality analysis</p><p><strong>final phase:</strong></p><p>- <strong>comment-analyzer: </strong>Verification of comment accuracy and maintainability</p><p>- <strong>code-simplifier: </strong>Suggestions for code simplification</p><p>Phases control execution order. Agents in the main phase run first, followed by agents in the final phase.</p><h3>How the Review Pipeline Works</h3><p>The hachimoku review pipeline consists of three stages: Selector, Review Agents, and Aggregator.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UbCFgmK1mbf1tIwp690q1w.png" /></figure><p>The <strong>Selector</strong> analyzes the changed content and selects which agents to run based on each agent’s applicability conditions (file patterns, content patterns). Rather than running every agent every time, only the agents relevant to the changes are selected.</p><p>Selected agents run <strong>in parallel by default</strong>.</p><p>The <strong>Aggregator</strong> consolidates all agent results, eliminates duplicate findings, and generates prioritized recommended actions.</p><h3>Installation and Quick Start</h3><p>With <a href="https://docs.anthropic.com/en/docs/claude-code/overview">Claude Code</a> and <a href="https://docs.astral.sh/uv/getting-started/installation/">uv</a> installed, run the following:</p><pre># Install<br>uv tool install git+https://github.com/drillan/hachimoku.git<br># Initialize your project<br>8moku init</pre><p>Running 8moku init creates a .hachimoku/ directory with default configuration files and a directory for custom agents.</p><h3>Three Review Modes</h3><pre># diff mode: Review changes on the current branch (no arguments)<br>8moku<br><br># PR mode: Review a GitHub PR (specify PR number)<br>8moku 42<br><br># file mode: Review specified files<br>8moku src/main.py tests/test_main.py</pre><p>The mode is determined automatically by the argument type: no argument for diff, a number for PR, and file paths for file mode.</p><h3>Listing Agents</h3><pre># List all agents<br>8moku agents<br># Show agent details<br>8moku agents code-reviewer</pre><h3>Model Selection Strategy</h3><p>hachimoku runs within Claude Code’s flat-rate plan, so no additional API charges apply.</p><p>Using lightweight Haiku for the Selector and Aggregator, and high-accuracy Opus for review agents, lets you balance cost and quality.</p><p>- <strong>Selector / Aggregator: </strong>→ Haiku — Lightweight and fast; selection and aggregation are simple tasks</p><p>- <strong>Review Agents:</strong> → Opus — High accuracy; deep code analysis required</p><p>- <strong>Lighter Reviews:</strong> → Sonnet — A middle ground between Opus and Haiku for speed vs. quality</p><p>You can configure models per agent in .hachimoku/config.toml:</p><pre># Default model (applies to all agents)<br>model = &quot;claudecode:claude-opus-4-6&quot;<br># Use a lightweight model for the selector<br>[selector]<br>model = &quot;claudecode:claude-haiku-4-5&quot;<br># Use a lightweight model for the aggregator<br>[aggregation]<br>model = &quot;claudecode:claude-haiku-4-5&quot;<br># Per-agent configuration<br>[agents.code-reviewer]<br>model = &quot;claudecode:claude-opus-4-6&quot;<br>[agents.comment-analyzer]<br>enabled = false  # Disable agents you don&#39;t need</pre><h3>Defining Custom Agents with TOML</h3><p>The defining feature of hachimoku is the ability to define agents in TOML files. You can add project-specific review perspectives without touching any code.</p><h3>Agent Definition Structure</h3><p>Here’s the structure, using the built-in code-reviewer agent definition (abbreviated) as an example:</p><pre>name = &quot;code-reviewer&quot;<br>description = &quot;Overall review for code quality, bugs, and best practices&quot;<br>model = &quot;claudecode:claude-opus-4-6&quot;<br>output_schema = &quot;scored_issues&quot;<br>phase = &quot;main&quot;<br>allowed_tools = [&quot;git_read&quot;, &quot;gh_read&quot;, &quot;file_read&quot;]<br>system_prompt = &quot;&quot;&quot;<br>You are code-reviewer, an expert code quality analyst.<br>...<br>&quot;&quot;&quot;<br>[applicability]<br>always = true</pre><p>- name: Agent name (lowercase alphanumeric and hyphens)</p><p>- description: Brief description of the agent’s role</p><p>- model: LLM model to use</p><p>- output_schema: Output schema name (`scored_issues` for scored findings, `severity_classified` for severity-based classification, etc.)</p><p>- phase: Execution phase. Runs in order: `early` → `main` → `final`</p><p>- allowed_tools: Permitted tool categories (`git_read`, `gh_read`, `file_read`, `web_fetch`)</p><p>- system_prompt: System prompt for the agent</p><p>- [applicability]: Conditions for when to run (`always`, `file_patterns`, `content_patterns`)</p><p>The [applicability] section lets you specify whether an agent should always run or only under certain conditions:</p><pre># Always run<br>[applicability]<br>always = true<br># Run only when Python files are changed<br>[applicability]<br>file_patterns = [&quot;*.py&quot;, &quot;*.pyi&quot;]<br># Run only when changes contain specific patterns<br>[applicability]<br>content_patterns = [&quot;import asyncio&quot;, &quot;async def&quot;]</pre><h3>Custom Agent Example: scope-drift-detector</h3><p>Here’s a custom agent actually used in a real project: scope-drift-detector.</p><p>This agent was created after encountering a real problem where an AI-driven refactoring PR inadvertently modified built-in agent TOML files outside its intended scope. It verifies that PR changes stay within the scope of the related Issue, and if out-of-scope changes are found, it suggests reverting them or splitting them into a separate PR.</p><pre>name = &quot;scope-drift-detector&quot;<br>description = &quot;Detects out-of-scope changes — verifies PR changes stay within the related Issue&#39;s scope&quot;<br>model = &quot;claudecode:claude-opus-4-6&quot;<br>output_schema = &quot;severity_classified&quot;<br>phase = &quot;main&quot;<br>allowed_tools = [&quot;git_read&quot;, &quot;gh_read&quot;, &quot;file_read&quot;]<br>system_prompt = &quot;&quot;&quot;<br>You are scope-drift-detector, an expert at detecting changes that drift outside the scope of a related Issue.## Purpose<br>Detect whether a PR contains changes outside the scope of the related Issue.<br>Out-of-scope changes carry risks of unintended side effects, regressions, and inconsistencies.<br>## Review Procedure<br>1. Read the selector analysis context (change_intent and issue_context) from the user message to understand the PR&#39;s intended scope.<br>2. If there is no issue_context, skip scope analysis and only perform project invariant checks.<br>3. Get the list of changed files with run_git([&quot;diff&quot;, &quot;--name-only&quot;]).<br>4. For each changed file, check the specific changes with run_git([&quot;diff&quot;, &quot;--&quot;, &quot;&lt;path&gt;&quot;]).<br>5. Determine whether each file&#39;s changes fall within the Issue&#39;s scope.<br>6. Perform project invariant checks on all changed files regardless of scope.<br>## Scope Determination Criteria<br>Considered &quot;in scope&quot;:<br>- Files/modules/areas explicitly mentioned in change_intent or issue_context<br>- Direct requirements of the feature described in the Issue<br>- Necessary follow-up changes from in-scope modifications (e.g., import updates after module moves, test fixes broken by in-scope changes)<br>Considered &quot;out of scope&quot;:<br>- Changes to files unrelated to the Issue&#39;s purpose that aren&#39;t a necessary consequence of in-scope changes<br>- &quot;While I&#39;m at it&quot; improvements to code the Issue doesn&#39;t target (refactoring, style fixes, comment updates)<br>- Changes to default values, configurations, or definitions not requested by the Issue<br>When scope is ambiguous, use read_file to check the full context of the changed file before making a determination.<br>Err on the side of not reporting - only report when confident that a change is out of scope.<br>## Severity Classification<br>Classify each finding into the appropriate list:<br>- critical_issues: Project invariant violations<br>- important_issues: Out-of-scope changes that alter behavior or values<br>- suggestion_issues: Changes that appear unrelated to the Issue but can&#39;t be definitively called out of scope<br>- nitpick_issues: Minor out-of-scope changes with no behavioral impact (whitespace, formatting)<br>## Confidence Filtering<br>Assign an internal confidence score (0-100) to each finding.<br>Only report findings with confidence 80 or above.<br>## Quality Principles<br>- Only report genuine scope drift or invariant violations<br>- Read issue_context carefully. Many changes that appear unrelated at first glance are actually required by the Issue<br>- Prefer a small number of high-confidence findings over a large number of uncertain ones<br>&quot;&quot;&quot;<br>[applicability]<br>always = true<br></pre><p>To add a custom agent, simply place a TOML file in the .hachimoku/agents/ directory. hachimoku automatically loads it at startup.</p><h3>Review History</h3><p>hachimoku automatically saves review results in JSONL format to the .hachimoku/reviews/ directory. In PR mode, results are saved as pr-&lt;number&gt;.jsonl. In diff mode, they are saved to a file based on the branch name.</p><p>Each record includes the review mode, commit hash, branch name, review timestamp, each agent’s findings (severity, description, file location, suggested fix), and a summary (total issue count, maximum severity).</p><p>Here’s an example review history entry for PR #227:</p><pre>{<br>  &quot;review_mode&quot;: &quot;pr&quot;,<br>  &quot;commit_hash&quot;: &quot;750369df...&quot;,<br>  &quot;pr_number&quot;: 227,<br>  &quot;branch_name&quot;: &quot;docs/225-web-fetch-docs-update&quot;,<br>  &quot;reviewed_at&quot;: &quot;2026-02-18T01:11:31.845505Z&quot;,<br>  &quot;results&quot;: [<br>    {<br>      &quot;status&quot;: &quot;success&quot;,<br>      &quot;agent_name&quot;: &quot;code-reviewer&quot;,<br>      &quot;issues&quot;: [<br>        {<br>          &quot;severity&quot;: &quot;Important&quot;,<br>          &quot;description&quot;: &quot;The SelectorDefinition allowed_tools default description in docs/agents.md...&quot;,<br>          &quot;location&quot;: {<br>            &quot;file_path&quot;: &quot;docs/agents.md&quot;,<br>            &quot;line_number&quot;: 41<br>          },<br>          &quot;suggestion&quot;: &quot;Revert &#39;4 categories&#39; back to the original &#39;3 categories&#39;...&quot;<br>        }<br>      ]<br>    }<br>  ],<br>  &quot;summary&quot;: {<br>    &quot;total_issues&quot;: 1,<br>    &quot;max_severity&quot;: &quot;Important&quot;<br>  }<br>}</pre><p>Since data is accumulated in a structured format, it can be used for trend analysis and review quality tracking.</p><h3>Integration with Issue Workflow</h3><p>hachimoku can also be combined with Issue Workflow. You can use hachimoku as the review step in Issue Workflow’s development flow for more flexible reviews. See the <a href="https://github.com/drillan/issue-workflow/blob/main/README.ja.md">Issue Workflow README</a> for details.</p><h3>Summary</h3><p>hachimoku is a multi-agent code review CLI where agents are defined in TOML.</p><ul><li>Six built-in agents provide multi-perspective reviews out of the box</li><li>The Selector automatically chooses agents based on changes, and the Aggregator consolidates the results</li><li>TOML-based custom agents let you add project-specific review perspectives without code changes</li><li>Runs within Claude Code’s flat-rate plan</li></ul><p>For more details, see the <a href="https://drillan.github.io/hachimoku/en/">documentation</a> and <a href="https://github.com/drillan/hachimoku/blob/main/README.md">README</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eee27dbf65f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Knowledge Management for Efficient Development with Claude Code]]></title>
            <link>https://medium.com/@patraqushe/knowledge-management-for-efficient-development-with-claude-code-bdd88b22bd68?source=rss-39e7a82c8009------2</link>
            <guid isPermaLink="false">https://medium.com/p/bdd88b22bd68</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[practice]]></category>
            <dc:creator><![CDATA[driller]]></dc:creator>
            <pubDate>Tue, 17 Jun 2025 13:31:55 GMT</pubDate>
            <atom:updated>2025-06-17T13:34:08.341Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>Claude Code is a tool that allows direct command-line access to “Claude,” an AI assistant provided by Anthropic. However, as projects grow, establishing a system to effectively communicate past trials and errors, as well as design decisions, to Claude becomes crucial.</p><p>This article introduces a practical methodology for systematically accumulating and leveraging knowledge gained from development using Claude Code.</p><h3>Challenges in Knowledge Management</h3><p>When you start using Claude Code, you may encounter the following challenges:</p><ul><li>The need to explain the same problem repeatedly.</li><li>Claude’s lack of understanding regarding the reasons behind past design decisions.</li><li>The effort required to convey project-specific constraints and requirements each time.</li><li>The scattering of knowledge gained from debugging sessions.</li></ul><p>To address these challenges, it is essential to build a structured knowledge management system.</p><h3>Proposed Knowledge Management System</h3><h3>Basic Directory Structure</h3><pre>project-root/<br>├── CLAUDE.md                    # Main configuration for Claude Code<br>└── .claude/<br>    ├── context.md              # Project background and constraints<br>    ├── project-knowledge.md    # Technical knowledge and patterns<br>    ├── project-improvements.md # Improvement history and lessons learned<br>    ├── common-patterns.md      # Frequently used command patterns<br>    ├── debug-log.md           # Important debug records<br>    └── debug/                 # Temporary debug files<br>        ├── sessions/          # Session-specific logs<br>        ├── temp-logs/         # Temporary files during work<br>        └── archive/           # Storage for resolved issues</pre><h3>Role and Operation of Each File</h3><h4>1. CLAUDE.md - Main Configuration File</h4><p>This file describes the project overview and instructions for Claude:</p><pre># Project Overview<br>This project is...<br><br>## Knowledge Management System<br>This project systematically manages knowledge using the following files:<br>### `.claude/context.md`<br>- Project background, objectives, and constraints<br>- Reasons for technology stack selection<br>- Business requirements and technical constraints<br>### `.claude/project-knowledge.md`<br>- Implementation patterns and design decision knowledge<br>- Reasons for architectural choices<br>- Patterns to avoid and anti-patterns<br>### `.claude/project-improvements.md`<br>- Records of past trials and errors<br>- Failed implementations and their causes<br>- Improvement processes and results<br>### `.claude/common-patterns.md`<br>- Frequently used command patterns<br>- Standard implementation templates<br>**Important**: When making new implementations or significant decisions, update the relevant files.</pre><h4>2. .claude/context.md - Project Context</h4><pre># Project Context<br><br>## Overview<br>- Project Name: EC Site Management System<br>- Technology Stack: React + TypeScript + Node.js<br>- Goal: Build an easy-to-use EC site for small and medium-sized businesses<br><br>## Constraints<br>- Maintain compatibility with existing PostgreSQL database<br>- Response time within 200ms<br>- No IE11 support required (modern browsers only)<br><br>## Reasons for Technology Selection<br>- React: High team proficiency<br>- TypeScript: Emphasis on type safety and maintainability<br>- Tailwind CSS: For rapid UI development</pre><h4>3. .claude/project-knowledge.md - Technical Knowledge</h4><pre># Project Knowledge Base<br><br>## Architecture Decisions<br>### Authentication System<br>- Adopted JWT + Refresh Token pattern<br>- Reason: Balance of scalability and security<br>- Learning: Initially session-based, changed with a view to microservices<br><br>## Implementation Patterns<br>### Error Handling<br>- Use a common ErrorBoundary component<br>- API errors handled uniformly in `src/utils/errorHandler.ts`<br>- Learning: Individual handling in each component was repetitive and inefficient<br><br>## Patterns to Avoid<br>- Excessive `useState` division (performance issues)<br>- Deep nesting of CSS-in-JS (increased bundle size)<br><br>## Library Selection<br>- State Management: Zustand (lighter than Redux, lower learning curve)<br>- Testing: Jest + React Testing Library (emphasis on team proficiency)</pre><h4>4. .claude/project-improvements.md - Improvement History</h4><pre># Improvement History<br><br>## 2024-12-15: Page Display Speed Improvement<br>**Problem**: Initial display of the product list page took over 3 seconds<br><br>**Trial and Error**:<br>- ❌ Image lazy loading → Limited effect<br>- ❌ SQL query optimization → About 10% improvement<br>- ✅ Gradual loading of product data → 60% improvement<br><br>**Final Solution**:<br>1. Initial display of only 20 items<br>2. Additional loading with infinite scroll<br>3. WebP conversion and CDN delivery of images<br><br>**Lesson Learned**: From a UX perspective, gradual loading is more effective than displaying all items at once.</pre><h4>5. .claude/common-patterns.md - Frequently Used Patterns</h4><pre># Frequently Used Patterns<br><br>## Component Generation<br>```bash<br># Create a new React component<br>claude create component [ComponentName] with TypeScript and styled-componentsma<br>```</pre><h3>Debugging Support</h3><pre># Analyze error log and suggest fixes<br>claude analyze error log and suggest fixes for [error-description]</pre><h3>Managing Debug Information</h3><h4>Directory Structure and Operating Rules</h4><pre>.claude/debug/<br>├── sessions/          # Temporary logs per session<br>├── temp-logs/         # Temporary files during work<br>└── archive/           # Storage for resolved issues</pre><h3>Persistence Rules</h3><p>Problems meeting the following conditions should be recorded in debug-log.md:</p><ul><li>Problems that took more than 30 minutes to resolve</li><li>Problems with a high probability of recurrence</li><li>Knowledge that should be shared across the entire team</li></ul><h3>Recording Format</h3><pre>## [YYYY-MM-DD] Problem Summary<br>**Symptoms**: Error messages or abnormal behavior<br>**Environment**: OS, Node.js, Browser version<br>**Reproduction Steps**: Specific operating procedures<br>**Trial and Error**: Methods attempted and their results<br>**Final Solution**: Reliably working resolution steps<br>**Root Cause**: The fundamental cause of the problem<br>**Preventive Measures**: Measures to avoid the same problem</pre><h3>Tips for Effective Prompts</h3><h3>1. Clearly Provide Context</h3><pre>This project is a React + TypeScript web application.<br>Currently, I am implementing user authentication and using JWT tokens.<br>Considering the constraints in .claude/context.md, please implement the following features...</pre><h3>2. Specify Concrete Deliverables</h3><pre>Please create an API endpoint with the following requirements:<br>- RESTful design<br>- Include TypeScript type definitions<br>- Include error handling<br>- Generate test code<br>- Follow the patterns in .claude/project-knowledge.md</pre><h3>3. Leverage Past Knowledge</h3><pre>Based on the lessons learned from &quot;Page Display Speed Improvement&quot; in .claude/project-improvements.md,<br>please implement the product search function.</pre><h3>Updates via Commands</h3><p>By registering <a href="https://docs.anthropic.com/ja/docs/claude-code/tutorials#%E3%82%AB%E3%82%B9%E3%82%BF%E3%83%A0%E3%82%B9%E3%83%A9%E3%83%83%E3%82%B7%E3%83%A5%E3%82%B3%E3%83%9E%E3%83%B3%E3%83%89%E3%82%92%E4%BD%9C%E6%88%90%E3%81%99%E3%82%8B">custom slash commands</a>, you can update knowledge at any time.</p><pre>When new knowledge is gained, please update the following files:<br><br>- `.claude/context.md`: Background, constraints, technology selection reasons<br>- `.claude/project-knowledge.md`: Implementation patterns, design decisions<br>- `.claude/project-improvements.md`: Trials and errors, improvement records<br>- `.claude/common-patterns.md`: Standard implementations, command patterns</pre><p>After creating this command file, when new technical discoveries or design decisions occur during project work, simply executing the /project:learnings command will allow Claude Code to organize and record the knowledge in the appropriate files.</p><p><strong>Note</strong>: For examples of utilizing custom slash commands, please refer to the “Related Articles” at the end of this article.</p><h3>Actual Operational Effects</h3><p>The following effects have been observed in projects that adopted this system:</p><h3>Improved Development Efficiency</h3><ul><li>Reduced repetition of explanations</li><li>Decreased design inconsistencies during new feature development</li><li>Shortened debugging time</li></ul><h3>Improved Team Collaboration</h3><ul><li>Onboarding time for new members halved</li><li>Increased transparency of design decisions</li><li>Prevention of knowledge individualization</li></ul><h3>Improved Code Quality</h3><ul><li>Consistent code structure</li><li>Prevention of anti-pattern recurrence</li><li>Suppression of technical debt accumulation</li></ul><h3>Operational Considerations</h3><h3>1. Continuous Updates</h3><p>Knowledge files must be continuously updated as “living documents.” Make it a habit to update them immediately when important decisions or learnings occur.</p><h3>2. Information Organization</h3><p>Over time, information can become outdated or redundant. Regular reviews, perhaps once a month, are recommended to organize information.</p><h3>3. Team Sharing</h3><p>Don’t keep knowledge personal; sharing it across the entire team will lead to improved overall organizational development efficiency.</p><h3>Important Commands During Operation</h3><h3>Utilizing the /init Command</h3><p>To effectively operate this system, it is recommended to execute the /init command at the following timings:</p><pre>claude /init</pre><p><strong>When to Execute</strong>:</p><ul><li><strong>Initial Introduction</strong>: Creation of the .claude/ directory and initial setup</li><li><strong>After Major Knowledge File Updates</strong>: When significant changes are made to project-knowledge.md or context.md</li><li><strong>When New Members Join</strong>: To ensure the latest project context is accurately reflected</li><li><strong>After Regular Maintenance</strong>: After monthly knowledge organization or file structure changes</li></ul><p>Executing /init allows Claude Code to accurately grasp the latest project state and provide consistent suggestions.</p><h3>Conclusion</h3><p>To effectively utilize Claude Code, it is important to systematically manage project context and past knowledge, not just handle one-off questions.</p><p>By introducing the knowledge management system introduced in this article, you can expect:</p><ul><li>Significant improvement in development efficiency</li><li>Improved team collaboration</li><li>Enhanced code quality</li><li>Reduction of technical debt</li></ul><p>Claude Code is still a new tool, but by combining it with an appropriate knowledge management system, you can realize a next-generation development experience utilizing AI assistants.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bdd88b22bd68" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>