• Spec-Driven Flow: Before writing code with an agent, define a concise spec: problem statement, acceptance criteria, and what is out-of-scope.
• Verification-First Tasking: Give coding agents an explicit, checkable success condition (e.g., passing tests, lint clean) rather than a prescriptive step-by-step recipe. Verification loops matter more than prompt precision.
• Reusable Skills: Rather than prompting from scratch for complex workflows, leverage predefined agent skills (e.g., structured PR reviews, architecture audits). Curate and share these routines across the team to standardise agent behaviours.
• Tooling via MCP: Extend your assistant's capabilities using the Model Context Protocol (MCP). Use MCP servers to securely connect agents to internal databases, live documentation, and specialised APIs, eliminating the need to manually copy-paste context.
• Review Boundaries: For unfamiliar or high-stakes code, rely on tighter human review. For well-scoped, checkable work, rely on automated verification.
• Context is Key: Maintain an AGENTS.md file at the root of the repo with project context, schema, workflow rules, and critical gotchas.