Optimizing Claude Code Workflows

The Breakthrough

The most effective way to improve agent performance is to treat your system setup as technical debt, deleting any instructions or constraints that the current model generation can now handle autonomously.

What Actually Worked

• Prune the system prompt: Use the doctor command to automatically trim claude.md and custom skills. If the model can infer the task or command without your explicit instruction, delete the instruction.
• Implement strict failure-based evals: Define success criteria that can explicitly fail, such as pixel-by-pixel screenshot comparisons or automated test suites. Without a hard stop condition, agents will stall or hallucinate progress.
• Shift from imperative to outcome-based prompting: Instead of detailing how to perform a task, describe the desired state and constraints. Providing inspiration images or rigid step-by-step guides causes the model to copy the input rather than reasoning through the problem.
• Optimize context loading: Move instructions into claude.md only if they are required for every session. For occasional or complex tasks, move instructions into separate skill files that are only read when needed, reducing unnecessary token usage.
• Adopt cloud-based routines: Replace complex, expensive dynamic agent workflows with simple, recurring cloud jobs that handle maintenance tasks like cleaning up unused code, running tests, or merging duplicate features while the user is offline.

Context

As models like Claude 3.5 Sonnet and Opus 5 have become more capable, many developers continue to use "handholding" techniques—such as verbose system prompts and rigid command sequences—that were necessary for older, less intelligent models. These legacy setups consume context window space and often conflict with the model's native reasoning capabilities. The goal is to move toward a lean configuration where the agent is given the objective and the boundaries, rather than a script to follow.

Notable Quotes

• "The rule: would Claude have worked this out on its own? If yes, delete it."
• "An eval is a fixed check that decides when work is done, and it has to be strict enough to fail."
• "Anything concrete you hand it, it treats as something to copy instead of something to learn from."