{"slug": "claude-md-agents-md-should-be-a-bootloader-not-a-knowledge-base", "title": "claude.md/agents.md Should Be a Bootloader, Not a Knowledge Base", "summary": "A developer argues that AGENTS.md should function as a bootloader rather than a knowledge base, advocating for staged context delivery via the Model Context Protocol (MCP). The developer explains that MCP can serve as a pluggable context layer, enabling structured packages like invariants, skills, workflows, and resolvers to keep AI models focused and scalable.", "body_md": "In my previous post, I wrote that MCP may be more useful as a context distribution layer than as a simple RPC mechanism.\n\nThe discussion that followed made the idea clearer.\n\nThe real point is not “how to use MCP.”\n\nThe real point is:\n\nHow should we give context to AI systems in stages?\n\nMCP is useful because it gives us a clean transport for that staged context.\n\nIt can expose documents.\n\nIt can expose resolvers.\n\nIt can expose workflows.\n\nIt can expose skills.\n\nIt can expose operating contracts.\n\nThat means MCP is not only a tool-calling interface.\n\nIt can become a pluggable context layer for AI-assisted work.\n\nMany AI coding setups rely on local instruction files.\n\nExamples include:\n\nAt first, this works well.\n\nYou write a few rules.\n\nThen you add coding conventions.\n\nThen architectural constraints.\n\nThen domain knowledge.\n\nThen workflow notes.\n\nThen testing rules.\n\nThen risk warnings.\n\nThen things the AI should never do.\n\nThen things the AI should always check.\n\nEventually, the instruction file becomes too large.\n\nThen a new ritual begins:\n\ncompress the context so the AI can use it.\n\nThis becomes part of the daily cost of using AI.\n\nPeople maintain prompts.\n\nPeople compress documents.\n\nPeople remove old rules.\n\nPeople rewrite context.\n\nPeople tune instructions for each client.\n\nThe result is fragile.\n\nThe AI output depends on how well each user maintains their local context.\n\nThat is not a scalable team system.\n\nI think AGENTS.md should have a smaller role.\n\nAGENTS.md should not contain all domain knowledge.\n\nIt should not contain every workflow.\n\nIt should not contain every skill.\n\nIt should not become a compressed version of the organization.\n\nInstead:\n\nAGENTS.md should be a bootloader.\n\nIts job should be simple:\n\nThat is all.\n\nThe detailed knowledge should live elsewhere.\n\nThe startup file should point to the context system.\n\nIt should not become the context system.\n\nOnce context is provided through MCP, the architecture changes.\n\nBefore MCP:\n\nThe user carries the context locally.\n\nAfter MCP:\n\nThe MCP server provides the context.\n\nThis is a big difference.\n\nA user no longer needs a full local checkout of the governance repository.\n\nA user no longer needs to maintain a giant prompt.\n\nA user no longer needs to manually copy the latest domain rules.\n\nThe client only needs:\n\nThe context itself becomes pluggable.\n\nProject A can use one MCP context server.\n\nProject B can use another.\n\nA domain team can provide its own skill catalog.\n\nA governance team can maintain shared operating contracts.\n\nThe AI client becomes lighter.\n\nThe domain context becomes centrally maintained.\n\nWhen people think about giving context to AI, they often imagine sending everything at once.\n\nAll documents.\n\nAll rules.\n\nAll constraints.\n\nAll domain knowledge.\n\nAll examples.\n\nAll workflows.\n\nThis creates a new problem.\n\nThe context becomes too large.\n\nImportant rules become diluted.\n\nThe model receives information that is not needed for the current task.\n\nStable rules and volatile state get mixed together.\n\nThe AI may follow the wrong document, the wrong workflow, or the wrong level of detail.\n\nMore context does not always mean better output.\n\nSometimes, too much context makes the AI less reliable.\n\nThis is why context should not be delivered as a single dump.\n\nIt should be delivered as structured packages.\n\nFor example:\n\nEach package should have a clear purpose.\n\nStartup context should only contain invariants.\n\nA Skill should contain the knowledge and procedure for one kind of work.\n\nA workflow should define the expected sequence of work.\n\nA resolver should fetch authoritative documents when needed.\n\nRuntime tools should fetch volatile state only when needed.\n\nThis keeps the model focused.\n\nThe AI does not need the entire organization in its context window.\n\nIt needs the right context at the right stage of work.\n\nThis is where MCP becomes useful.\n\nMCP gives us named entry points for context.\n\nInstead of pushing one huge prompt into the model, the client can ask for:\n\nThat makes context staged, explicit, and easier to reason about.\n\nThe goal is not to maximize context size.\n\nThe goal is to control context shape.\n\nIf MCP can deliver documents, it can deliver more than documents.\n\nA document is context.\n\nA Skill is also context, but with a stronger structure.\n\nA good Skill does not only say:\n\nHere is some information.\n\nA good Skill says:\n\nHere is how this work should be done.\n\nA domain Skill can include:\n\nThis is much more valuable than simply retrieving document chunks.\n\nIf documents distribute knowledge, Skills distribute work quality.\n\nThat is the key point.\n\nMany AI Skills today are command-oriented.\n\nThey are useful, but they are often too low-level.\n\nFor example:\n\nThis looks like automation.\n\nBut in practice, it often becomes micromanagement.\n\nThe human still has to decide:\n\nThe AI executes small operations.\n\nThe human manages the workflow.\n\nThat does not create a large productivity gain.\n\nIt only changes the interface.\n\nThe user is still steering every step.\n\nThe hard part of professional work is not always execution.\n\nThe hard part is judgment.\n\nFor software work, the important questions are often:\n\nGeneric command bundles do not answer these questions.\n\nThey automate operations, not judgment.\n\nThat is why command-level Skills can improve convenience without improving team-level output quality.\n\nThey reduce keystrokes.\n\nThey do not necessarily reduce variance.\n\nA better Skill boundary is not a command.\n\nA better Skill boundary is a business-level work unit.\n\nExamples:\n\nThese are not single commands.\n\nThey are units of work.\n\nBut there is an important point here:\n\nDomain knowledge alone is not enough.\n\nA repository may contain many documents.\n\nA team may have many rules.\n\nA project may have many constraints.\n\nAn organization may have a large amount of accumulated knowledge.\n\nBut giving all of that knowledge to the model does not automatically improve the work.\n\nThe model does not need all domain knowledge.\n\nIt needs the knowledge that is necessary for the current work.\n\nAnd it needs that knowledge at the right moment.\n\nThat is why a domain Skill should not only contain instructions.\n\nA domain Skill should define:\n\nIn other words, a Skill is not just a procedure.\n\nA Skill is a work unit with controlled access to domain knowledge.\n\nThis is where MCP becomes useful as a context distribution layer.\n\nThe Skill does not need to embed every document directly.\n\nThe startup context does not need to preload the entire domain.\n\nThe client does not need to maintain a giant local prompt.\n\nInstead, MCP can provide the Skill and the knowledge access path.\n\nThe AI can load:\n\nThis makes domain knowledge usable.\n\nThe value is not in storing knowledge.\n\nThe value is in delivering the right knowledge for the right work unit.\n\nA senior engineer can define the Skill.\n\nThe Skill can point to the required domain knowledge.\n\nThe team can use the Skill through MCP.\n\nThe AI can follow the same rules each time.\n\nThe output becomes more consistent because the work unit, the knowledge access path, and the closure criteria are distributed together.\n\nThat is why domain Skills should be business-level units.\n\nThey are not command bundles.\n\nThey are packaged work contexts.\n\nIf Skills are business-level units, users should not have to manually pick every command.\n\nThe user should describe the work intent.\n\nThe system should route that intent to the right Skill.\n\nThat is why semantic routing matters.\n\nCommand routing says:\n\nWhich tool should I call?\n\nSemantic routing says:\n\nWhat kind of work is this?\n\nThat difference matters.\n\nIf the user must manually choose every command, the workflow stays at the micromanagement level.\n\nIf the system can route work intent to a domain Skill, the user can delegate at a higher level.\n\nThe Skill then carries the domain rules, references, unknown handling, and closure criteria.\n\nThis is closer to real delegation.\n\nWith MCP and semantic routing together, the model becomes different.\n\nThe user does not maintain a giant local prompt.\n\nThe user does not manually select every low-level command.\n\nThe user does not need a local copy of every governance document.\n\nInstead:\n\nThis is not just tool calling.\n\nThis is staged context delivery.\n\nThis is the layer that has been missing.\n\nIndividual AI use depends on personal prompt skill.\n\nGeneric Skills automate commands.\n\nRAG retrieves likely relevant knowledge.\n\nRPC lets the AI call tools.\n\nBut teams need something else.\n\nTeams need a way to distribute:\n\nThat is what domain Skills can provide.\n\nAnd MCP makes those Skills pluggable.\n\nThe main idea is simple:\n\nAGENTS.md should be a bootloader, not a knowledge base.\n\nAnd:\n\nMCP should make domain context and domain Skills pluggable.\n\nThis avoids the old pattern where every user maintains a growing local prompt.\n\nIt also avoids the trap of treating Skills as command bundles.\n\nFor team-level AI work, the goal is not to automate more commands.\n\nThe goal is to reduce quality variance.\n\nGeneric Skills automate operations.\n\nDomain Skills distribute judgment.\n\nThat is why I think MCP becomes most valuable when used for staged context delivery and domain Skill distribution.\n\nNot just RPC.\n\nNot just RAG.\n\nNot agent-to-agent coordination.\n\nA pluggable context layer for consistent AI-assisted work.", "url": "https://wpnews.pro/news/claude-md-agents-md-should-be-a-bootloader-not-a-knowledge-base", "canonical_source": "https://dev.to/synthaicode_commander/claudemdagentsmd-should-be-a-bootloader-not-a-knowledge-base-1lem", "published_at": "2026-06-26 08:57:51+00:00", "updated_at": "2026-06-26 09:34:02.239702+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "developer-tools"], "entities": ["MCP", "AGENTS.md", "CLAUDE.md"], "alternates": {"html": "https://wpnews.pro/news/claude-md-agents-md-should-be-a-bootloader-not-a-knowledge-base", "markdown": "https://wpnews.pro/news/claude-md-agents-md-should-be-a-bootloader-not-a-knowledge-base.md", "text": "https://wpnews.pro/news/claude-md-agents-md-should-be-a-bootloader-not-a-knowledge-base.txt", "jsonld": "https://wpnews.pro/news/claude-md-agents-md-should-be-a-bootloader-not-a-knowledge-base.jsonld"}}