PostHog Engineer's Claude Code Agent Workflow Practice: A Complete Solution from Code Review to Automated Deployment

PostHog engineer charlesvien has open-sourced his personal Claude Code agent workflow configuration repository (agents), a complete AI-assisted development workflow system specifically designed for the PostHog development team. Through its skill system (Skills), agent configuration (Agents), and Hooks mechanism, combined with structured prompts and automated tools, this system provides end-to-end support from code writing, testing, review to deployment, aiming to improve the development efficiency of large-scale projects and upgrade Claude Code into a true development partner.

The repository named "agents" is not just a collection of scripts but a complete AI-assisted development workflow system. Its design philosophy revolves around the Graphite stacked PR workflow and PostHog code ecosystem, aiming to upgrade Claude Code from a simple code assistant to a true development partner. The core idea of the project is to decompose development work into reusable "Skills" and specialized "Agents", and through well-designed prompt engineering and environment configuration, achieve end-to-end automated support from code writing, testing, review to deployment.

Ship (Code Delivery) Skills

ccheckpoint - Workflow checkpoint management: Create checkpoints bound to Git SHA, support verifying status or listing historical checkpoints, reducing the risk of experimental development.

cwrap - Parallel quality gate: Execute type checking, code linting, building, and testing tasks in parallel; automatically enter a repair loop when errors are found, shortening the feedback cycle.

cpr - Intelligent PR generation: Automatically generate PR titles and descriptions that comply with team norms based on uncommitted changes or branch differences, reducing documentation time.

crebase - Intelligent rebase assistant: Automatically rebase the current branch onto the parent branch and handle conflicts, saving time on manual conflict resolution.

creview - Automated code review: Analyze branch differences, categorize issues by severity and provide adjudication suggestions, serving as a pre-review filter.

Write (Code Writing) Skills

ctest - Co-located test generation: Generate test code in the same location as the specified source file, keeping the project structure clear.

crefactor - Parallel code refactoring: Execute code transformations across multiple files in parallel, ensuring consistency and completeness of refactoring.

Explore (Code Exploration) Skills

ctrace - tRPC call end-to-end tracing: Trace tRPC calls to help understand inter-service call relationships and data flow.

cdebug - Intelligent troubleshooting: Automatically categorize issues, locate root causes, and provide repair suggestions, systematizing the debugging process.

clogs - Session log location: Quickly find specific session log files for easy review and auditing.

The project configures two dedicated agents for different task scenarios:

code-reviewer (Opus Model)

Deep code review agent using the Claude Opus model to enhance reasoning capabilities. Review scope includes:

• Bug detection: Identify potential logical errors and edge cases
• Security audit: Check for common security vulnerabilities and sensitive information leakage risks
• Performance analysis: Identify code patterns that may cause performance bottlenecks
• TypeScript strictness check: Ensure code complies with strict type safety standards

refactor-cleaner (Sonnet Model)

Code cleaning dedicated agent using the Sonnet model to balance cost and performance. It discovers and safely removes unused code, dependencies, and exports through tools like knip, depcheck, and ts-prune, reducing technical debt.

Modern Command-Line Toolchain

The project recommends modern command-line tools to replace traditional Unix tools:

Tool	Replacement Target	Advantage
ripgrep (rg)	grep	Faster search speed, automatically respects .gitignore
fd	find	More concise syntax, better performance
bat	cat	Syntax highlighting and line number display
tree	ls -R	Visual directory structure display
delta	diff	Syntax-highlighted Git diff comparison
fzf	Manual search	Fuzzy search for files, branches, and history records

Environment Variable Optimization

Includes Claude Code environment variable optimization suggestions: Disable 1M context limit, disable adaptive thinking mode, disable automatic memory function, etc., to achieve more predictable behavior.

Hooks Mechanism

Configure the PreToolUse hook to monitor tool call counts: When tool calls exceed 50 times in a session, it is recommended to execute the /compact command to compress the context, then remind every 25 times afterward, managing the long session context window to prevent performance degradation.

This workflow system is most suitable for the following scenarios:

1. Large team collaboration: Standardized skills and agent configurations establish consistent development practices
2. Graphite workflow users: Skills are designed specifically for stacked PRs and work best with the Graphite toolchain
3. TypeScript/Node.js projects: Skills are optimized around this tech stack
4. Teams pursuing code quality: Automated review and test generation maintain high standards

Recommendation: Teams looking to introduce AI-assisted development can customize based on this project (a production-validated reference implementation) according to their own needs, and gradually build an agent ecosystem suitable for their team.

charlesvien's agents repository demonstrates a typical path for AI-assisted development from "toy" to "production tool". Through systematic prompt engineering, toolchain integration, and workflow design, Claude Code can transform from a conversational AI into a true development partner. With the accumulation and sharing of similar practices, we look forward to more mature AI-assisted development models emerging in the industry.