While `claude.md` files can guide AI behavior, they aren't always adhered to. Use Claude Code's "session start hooks" instead. They guarantee that critical context like goals, tasks, and past mistakes is injected into every new chat, making the AI more reliable.

Instead of manually providing context in each prompt, use Claude Code's 'append system prompt' command. This preloads crucial information, like architectural diagrams, at the start of a session, leading to faster and more accurate AI responses without repeated file reads.

Putting all instructions in a single `claude.md` file is inefficient. Instead, use the main file to act as a router, containing only high-level instructions on where to find specific knowledge (e.g., in `marketing_rules.md`). This keeps prompts efficient and scalable.

Developers found Claude Opus 5 broke existing workflows because Anthropic removed 80% of its internal system prompt. This architectural shift means older, detailed prompts now conflict with the new model's design. This highlights a significant, recurring migration cost for developers: each new model generation may require a complete rewrite of prompt libraries.

Contrary to past best practices, providing explicit examples within tool descriptions or system prompts can now degrade performance in advanced models. These models are imaginative enough to understand intention without being constrained by specific examples, which can negatively bias their output.

The easiest way to teach Claude Code is to instruct it: "Don't make this mistake again; add this to `claude.md`." Since this file is always included in the prompt context, it acts as a permanent, evolving set of instructions and guardrails for the AI.

Anthropic's Claude models are specifically trained on XML. By structuring system instructions using XML tags (e.g., <role>, <instructions>), you align with the model's training data. This provides better organization and can unlock additional functionality and more reliable outputs compared to using plain text prompts.

LLMs tend to amend instructions rather than replace them, leading to confusing and contradictory prompts over time. To maintain agent performance, periodically "purge" your markdown instruction files by rewriting them from scratch, ensuring they remain concise and accurate.

OpenAI found that removing repeated instructions from old prompts improved scores by 10-15% while cutting token usage by 66%. The complex rule lists built for older models now confuse systems like GPT-5.6, leading to worse and more expensive answers.