Claude Code is a powerhouse

The "Rule of Three" for Repo Skills

Before getting into the complex orchestration, there is one habit that separates the power users from the beginners: the Rule of Three. If you find yourself typing the same instruction or context three times across different sessions, stop and turn that instruction into a repo-specific skill.

In my current frontend workflow, we've already built skills for removing feature toggles and cleaning up warnings. Instead of explaining the project structure every time, the agent already knows exactly where to look and which tools to trigger. This is the single biggest leverage point for reducing prompt fatigue.

Serial vs. Parallel Agent Orchestration

When a task is too massive for a single conversation window, you have to decide how to orchestrate the workload. I've found two distinct patterns that work depending on whether you prioritize cost or speed.

• Serial Orchestration: This is essentially a resumable loop. You make one change, wait for the MR to merge, and then move to the next. I keep the state in a markdown file. This is the safest bet when you need a human-in-the-loop review for every single step. It's token-efficient because you aren't spawning multiple contexts.
• Parallel Orchestration: This is the "fan out" approach. You tell Claude to spawn multiple sub-agents to handle independent modules simultaneously. It's perfect for mindless, repetitive migrations or adding return types across a massive codebase. Each sub-agent works in its own worktree or commit, and the main orchestrator handles the final merge.

Pro Tip for Budgeting: Parallel execution can burn through your token quota terrifyingly fast if you aren't careful. The gold standard for a cost-effective AI workflow is: Orchestrate with Opus, execute sub-tasks with Sonnet. If you don't explicitly tell the orchestrator to use Sonnet for the workers, you might find yourself paying Opus prices for tasks that don't require that level of reasoning.

A Practical Tutorial for the Standard Flow

Despite the fancy orchestration, 90% of my day still follows a basic loop. If you're building an LLM agent workflow from scratch, stick to this sequence to avoid hallucinations:

1. Describe the goal in plan mode.
2. Iterate on that plan until the logic is airtight.
3. Trigger the execution.
4. Review the diffs.
5. Open the MR.

Essential Commands for Deep Dives

If you're doing a deep dive into a legacy codebase, the "deep research" capability is a lifesaver. It functions similarly to Perplexity, allowing the agent to synthesize information across the repo and external docs without you having to manually feed it every relevant file. It turns the agent from a coder into a researcher, which is where the real value lies when onboarding onto a complex project.