Skip to content

Instantly share code, notes, and snippets.

@arjunghosh
Last active March 22, 2026 23:58
Show Gist options
  • Select an option

  • Save arjunghosh/122b90e26e4df2e2d828baed69c39853 to your computer and use it in GitHub Desktop.

Select an option

Save arjunghosh/122b90e26e4df2e2d828baed69c39853 to your computer and use it in GitHub Desktop.

As a Technology Leader and CAIO, I have observed that the evolution of Agentic AI has shifted the focus from "prompt engineering" to "context engineering." The emergence of files like AGENTS.md, CLAUDE.md, and SKILLS.md represents a move toward standardized, repository-level governance for AI agents. The following table provides a structured comparison of these "Context Manifest" files used in modern AI-augmented development environments. Comparative Analysis: AI Context Manifest Files

Sl. No. File Name Primary Purpose Best Suited For Key Components
1 CLAUDE.md Environment & Task Governance Cursor / Claude Code users Build commands, test patterns, and style guide enforcement.
2 AGENTS.md Cross-Agent Protocol Multi-agent swarms (e.g., ProjectCodeX) High-level architectural intent, "Anti-patterns," and tribal knowledge.
3 SKILLS.md Tooling & Capability Mapping MCP (Model Context Protocol) Definitions of external tools, API schemas, and execution logic.
4 MEMORIES.md Persistent Learning Long-running projects Lessons learned from previous debugging sessions or refactors.
5 RULES.md Strict Linter-like Constraints Large, regulated teams Compliance requirements, security protocols, and "do-not-touch" zones.
Detailed Breakdown of Key Architectures
  1. CLAUDE.md: The Practical Workflow Guide Popularized by the Cursor IDE community, this file serves as a "local instruction set."
  • Intent: To prevent the LLM from hallucinating build commands or using deprecated libraries.
  • Content: Specific shell commands for testing (e.g., npm test -- --watchAll=false), project structure summaries, and preferred naming conventions (e.g., "Always use Functional Components").
  • Strategic Value: Reduces the "Discovery Phase" tokens spent by the agent when it first enters a repository.
  1. AGENTS.md: The Architectural Compass In the context of the SEOAG5 architectural approach, this file acts as the "North Star" for a collaborative swarm.
  • Intent: To bridge the gap between "what is in the code" and "why it is there."
  • Content: Business logic nuances that aren't obvious from the code, such as "We use this specific retry logic because the legacy API is unstable."
  • Strategic Value: As noted in recent research (referencing the Hacker News discussion), human-written AGENTS.md files provide a ~4% performance boost by offering non-obvious domain insights.
  1. SKILLS.md: The Capability Registry This is often used in Agentic Frameworks (like LangChain or custom MCP implementations) to define what an agent can do.
  • Intent: To map the "Action Space" of the agent.
  • Content: Descriptions of tools (e.g., "Database Explorer," "Jira Updater") and the parameters required to trigger them.
  • Strategic Value: It allows for a modular "Plug-and-Play" architecture where agents can be granted or revoked specific technical capabilities without changing their core logic.
  1. The "Memories" and "Rules" Paradigm
  • MEMORIES.md: Essential for FUTURE-RELEASE-PHASE2 (Predictive/Proactive Support). It captures the "History of Mistakes." If an agent spends 2 hours debugging a circular dependency, that "lesson" is stored here so the next agent doesn't repeat it.
  • RULES.md: Used for Foundational Security. It acts as a hard guardrail, ensuring the agent never attempts to bypass a Secure Prompting Gateway or modify sensitive /auth directories. External Resources & References | # | Resource Link | Description | |---|---|---| | 1 | Model Context Protocol (MCP) | The standard for connecting AI agents to external data/tools. | | 2 | Cursor Context Files Documentation | Official guide on how .cursorrules (a variant of these files) works. | | 3 | Agentic AI Coding Protocol | General industry standards for AI-assisted software development life cycles. | Strategic Recommendation For an enterprise-grade ecosystem like ProjectCodeX, I recommend a Hybrid Approach:
  • Use SKILLS.md for technical capability discovery (MCP-based).
  • Use AGENTS.md for architectural intent and tribal knowledge.
  • Use MEMORIES.md to ensure the agent swarm evolves with the project, effectively reducing technical debt over time. Note: Always ensure these files are kept under Version Control (Git) to maintain an Immutable Audit Trail, as per the [Global_Content_Protocol_v1.3].
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment