Modern AI models infer context from the codebase, making detailed `agent.md` files redundant. These files waste tokens on every interaction and are only necessary for highly specific, proprietary information that must always be present in the context.
Downloading skills from marketplaces is risky and ineffective. They are potential security vectors and lack the context of your specific workflows. For an agent to perform reliably, it must codify skills based on its direct interaction with your unique tasks, not a generic template.
Expect your AI agent's skills to fail initially. Treat each failure as a learning opportunity. Work with the agent to identify and fix the error, then instruct it to update the original skill file with the solution. This recursive process makes the skill more robust over time.
Instead of loading large context files on every turn, use "skills." The agent only sees a skill's name and description initially, loading the full instructions only when needed. This method, called progressive disclosure, drastically saves tokens and improves performance.
Top-tier language models are becoming commoditized in their excellence. The real differentiator in agent performance is now the 'harness'—the specific context, tools, and skills you provide. A minimalist, well-crafted harness on a good model will outperform a bloated setup on a great one.
Resist building complex, multi-agent systems from day one. Instead, start with a single agent and build its skills based on actual workflows. Add sub-agents only when a clear productivity need arises. This approach is more effective than scaling for what looks impressive.
Don't write agent skills from scratch. First, manually guide the agent through a workflow step-by-step. After a successful run, instruct the agent to review that conversation history and generate the skill from it. This provides the crucial context of what a successful outcome looks like.