Launch cloud code by typing '/initialize' to generate your cloud MD file and establish project settings automatically.

Use agents md as an extension beyond cloud code to develop and test sub-agents locally, improving iteration speed and context consistency.

Integrate Claude MD files to provide memory features in your agent workflow today while anticipating richer future memory capabilities.

Design a sub-agent that programmatically generates other sub-agents with built-in documentation and predefined configurations to streamline agent crea...Show more

Adopt Next.js directory conventions for agent folder structures to leverage familiar patterns and enhance meta-description management.

Build AI agents that generate and execute scripts automatically, expanding their utility beyond static responses into active task automation.

Specify JSON schema outputs directly in the agent markdown file to enforce structured response formats and simplify downstream parsing.

Define agent roles, system prompts, and output formats in an agent card structure to standardize agent descriptors and behaviors.

Configure memory and settings at the local project level to customize agent behavior and retain project-specific context.

Use sub-agents that spawn additional sub-agents to build scalable, layered AI workflows drawing on recursive architectures.

Organize prompts within a project folder to systematically manage context and variations, improving AI response consistency.

Reapplying the same prompt within a dedicated project folder to test consistency and performance across runs

Specifying and invoking particular AI agents directly through documentation to enforce structured usage patterns

Hiding internal reasoning traces in AI outputs to streamline responses or secure sensitive logic

Leveraging referencing and prescriptive language in AI documentation to specify and invoke AI agents directly

Defining roles such as academic papers specialist, code expert, trends analyst, and integration specialist to structure effective AI team workflows

Using Claude to break down tasks into sub-agents like academic trend analysis and code implementation for specialized AI workflows

Tom emphasizes defining sub-agents by narrowband context or specific outcomes rather than anthropomorphic roles to improve efficacy.

Cameron details using embeddings, semantic similarity, and task decomposition logic to automate sub-agent creation and composition.

Cameron suggests limiting sub-agents to around five to optimize AI system performance and avoid complexity.