At GleamHub, we completely overhauled the entire operational workflow for our corporate website (gleamhub.net) in early 2026, transitioning to an AI-driven system centered on Claude Code. In this article, we share our design philosophy and practical implementation details.
Why we chose Claude Code
When selecting an AI coding tool, we compared two broad approaches: IDE-integrated code completion tools versus terminal-based agentic tools. What ultimately drove our decision to adopt Claude Code was the granularity of task delegation.
IDE autocomplete suggests the immediate next line of code, but cannot handle operational-level directives like "publish this media article." Claude Code accepts high-level tasks in natural language via the terminal, executing file edits, Git workflows, and external API requests in an uninterrupted chain. That was the defining difference.
We compare Claude Code vs. Cursor in depth in this article.
High-level workflow architecture
Current operations for GH Media (our media section) run on an agent pipeline orchestrated by five custom slash commands:
| Command | Role |
|---|---|
/media-analytics | Aggregates page views, average engagement time, and traffic sources from GA4 to generate reports |
/media-strategy | Synthesizes search trends and historical performance metrics to propose content calendars |
/media-draft | Generates complete article drafts automatically based on provided keywords |
/media-seo | Scores frontmatter, article bodies, and internal links against SEO benchmarks |
/media-publish | Executes validation, PR creation, and social media post copy generation in a single pass |
These commands are defined as Markdown files under .claude/commands/ and loaded dynamically by Claude Code at runtime. Each command can be executed independently, or run sequentially to manage the entire lifecycle from initial ideation to publication.
Project memory via CLAUDE.md
Claude Code supports CLAUDE.md, which serves as the "constitution" for a project. Architecture rules, coding style guidelines, execution commands, and critical design constraints can be codified once, allowing Claude to load them across sessions and eliminating repetitive re-explanations.
Our CLAUDE.md includes information such as:
- Astro
trailingSlash: 'always'configuration and build commands - Content collection schema definitions (required frontmatter fields)
- Category slug mapping rules
- Deployment targets (Google Cloud Storage) and CI/CD pipelines
Thanks to this setup, instructions like "filter articles in draft: true" or "update categoryMapping.ts when introducing a new category" execute accurately from simple single-line prompts.
Ingesting external data via MCP integrations
Claude Code connects to third-party services through the Model Context Protocol (MCP). We implemented access to the GA4 Data API as a custom MCP server, enabling Claude Code to perform analytics queries directly.
# GA4 の先週トップ10記事を取得してトレンド分析
/media-analyticsWith a single command, it pulls metrics from GA4 and generates a Markdown report comparing performance across categories, week-over-week trends, and recommended action items. Tasks that previously took 2 to 3 hours of manual spreadsheet work each week now take mere minutes.
MCP offers extensive official integrations with core platforms such as Google Drive, Slack, and Notion, and creating custom MCP servers for proprietary tools incurs minimal development overhead. Its major operational strength lies in retrieving external data without polluting agent context.
Automating quality gates with Hooks
Claude Code's Hooks feature enables custom scripts to execute at key stages of an agent's lifecycle, such as after file modifications or prior to command execution.
We built the following quality checks into our Hooks:
- Frontmatter validation: Confirms that required fields
draft,title,excerpt, andfeaturedImageare present - Word count verification: Ensures the article body meets the minimum length requirement (3,000 characters)
- Internal link validation: Confirms that targeted internal link files actually exist
The key to effective Hooks is running them right before PR generation rather than interrupting every single write operation. This allows the agent to complete its operational flow uninterrupted, while quality gates firmly catch issues during final deliverable review.
Splitting large tasks with parallel agents
Claude Code's Parallel Agents feature allows large-scale refactoring jobs and batch processing across multiple articles to run concurrently.
In our own operations, we leveraged this to optimize internal links across 50 past articles. By grouping articles by category and assigning multiple subagents to work in parallel, we dramatically cut execution time compared to processing them sequentially with a single agent.
For more on designing agent orchestration, refer to our article on Automating Project Management with Claude Code.
Continuous improvement with Auto Memory
Introduced in the February 2026 update, Auto Memory automatically captures and accumulates project-specific context across sessions. Feedback such as "this article structure doesn't match our corporate voice" or "do not use this component" carries over cleanly into subsequent runs.
We maintain memory files under .claude/projects/ to record strategic principles (target personas, KGIs, and editorial guidelines). Because Claude references these directly, simple task-level requests consistently yield outputs aligned with our goals.
Changes before and after adoption
| Task | Before adoption | After adoption |
|---|---|---|
| Drafting an article | 2 to 3 hours | 20 to 30 minutes |
| SEO audit and revisions | 30 to 60 minutes | 5 to 10 minutes |
| OGP image generation | 30 minutes (manual work in Figma) | A single command (a few seconds) |
| PR creation and social media copy generation | 20 minutes | 3 minutes |
| GA4 weekly report | 2 to 3 hours | 5 to 10 minutes |
Beyond simply saving time, the biggest shift was that the very nature of the work changed. With less time spent executing tasks manually, we can now dedicate our time to decision-making, such as what to build and which directions to invest in.
Looking ahead
What we are currently exploring is fully automating weekly reporting using Claude Code's Scheduled Tasks feature. We aim to run an entirely hands-off workflow triggered every Monday morning that fetches GA4 data, generates a performance summary, and sends a notification to Slack.
AI agents are evolving from "usable tools" into genuine "team members." To make the most of this shift, we believe that design capabilities regarding what to delegate to agents—in other words, workflow design itself—becomes a core competitive advantage.
At GleamHub, we implement these AI-driven development workflows not only internally, but also for our client companies. If you would like to incorporate AI agents into your business operations or consult on improving development efficiency, please feel free to contact us.
