Agentic Coding

This GitHub repository documents advanced Claude Code practices directly aligned with your interests in agentic coding, efficient AI agent workflows, and best practices for solo SaaS development. Key relevant sections include detailed guidance on using agents, commands, and skills for structured workflows—particularly the Command → Agent → Skill orchestration pattern shown in the weather-orchestrator example. The repository emphasizes memory management through CLAUDE.md files (recommended under 200 lines), .claude/rules/ for domain-specific instructions, and context management techniques like using subagents to isolate work and prevent context pollution. For debugging workflows, it recommends taking screenshots to share with Claude, using MCP for browser debugging, and running terminal logs as background tasks. The prompt chaining approach is demonstrated through the development workflow pattern (Research → Plan → Execute → Review → Ship) and the tips section covers session management strategies including /compact for mid-task context cleanup, /rewind to backtrack from failed attempts, and using subagents specifically for context management. The repository also features 83 practical tips covering prompting strategies, context optimization, and advanced workflows like using /loop for recurring tasks and cross-model workflows for enhanced code review.

Claude Code trains you to steer agents through iterative workflows, while Codex trains you to dispatch them, fundamentally shaping how you approach AI-assisted work. The real challenge isn't getting machines to produce work—it's deciding when that work is good enough to accept, a decision that hinges on understanding agent failure modes like "theater" (convincing conversation without real understanding) and "completion theater" (finished runs that feel more done than they are). The author provides a practical framework including five shapes agent runs take, six pre-launch questions, four reusable prompts (Run Spec, steer-or-dispatch diagnostic, "is it real?" audit, and cross-check), and emphasizes that managing AI labor through these tools is creating a new skill set applicable across knowledge work beyond just coding.

OpenViking is a context database designed for AI agents that addresses memory management challenges through a filesystem-based paradigm. It unifies fragmented memories, resources, and skills into a structured virtual filesystem (viking:// protocol), eliminating the need for scattered vector databases. The system implements tiered context loading (L0 abstract, L1 overview, L2 details) to reduce token consumption, and uses directory recursive retrieval combining vector search with hierarchical positioning to improve retrieval accuracy. OpenViking provides visualized retrieval trajectories so developers can observe and debug the context retrieval process, and features automatic session management that extracts and iterates agent task memories over time. Benchmarks show significant improvements: with Claude Code, OpenViking achieved 80.32% accuracy (up from 57.21%), reduced latency by 58.45%, and cut token usage by 63.2%, while on multi-hop RAG tasks reaching 91% accuracy with minimal retrieval latency.

The author shares six months of practical experience using Claude Code for testing workflows. The testing pipeline has evolved into an automated chain where Claude reads Jira tickets and code to draft manual testing plans, then converts those into Postman collections and Cypress scenarios—but only after the author manually validates the plans against the actual application. This discipline of having AI draft and reality validate prevents plausible-but-wrong tests from entering the codebase. Beyond tests, the time freed up enabled building internal tools like a PostmanToCypressConverter CLI, a flakiness dashboard, and an internal Claude Code plugin marketplace with role-based naming conventions (qa-generate-tests, dev-review-pr, etc.). However, the major shift is that code review now dominates the workflow—reviewing AI-generated code is harder than writing it because you must reconstruct mental models at the speed AI produces them, catching subtle bugs like assertions that check the wrong thing while appearing correct. The author emphasizes that AI amplifies existing engineering judgment; without the skill to spot plausible-but-wrong code, the volume of generated code becomes dangerous rather than helpful.