Detailed Analysis
This Reddit thread surfaces a set of practical architecture questions that reflect how power users are adapting Claude Code to real-world, multi-tenant workflows rather than the simple single-project scenarios most documentation assumes. The poster runs a solo marketing agency managing roughly fifteen clients within a single repository, each with its own CLAUDE.md file containing locked MCP account IDs, contact information, and account-specific rules. The core technical question is whether to always launch Claude Code sessions from the repository root versus changing directory into a client subfolder, a decision that initially seems like a simple token-optimization tradeoff but turns out to have deeper implications for how Claude Code's memory, settings, and skills systems behave.
The poster's key discovery is that nested CLAUDE.md files load lazily, only being pulled into context when a file within that subtree is actually touched, which means starting in a subfolder doesn't guarantee early loading of the right account context anyway. Meanwhile, memory, settings, and skills in Claude Code are keyed to the launch directory, so inconsistently starting sessions from different subfolders fragments state across the workspace. This is a subtle but important finding for anyone managing multiple isolated contexts in one repo: the tool's session-scoping behavior means directory choice affects more than just which files get read, it affects the persistence layer itself. The poster's workaround, always launching from root and adding an explicit root-level rule instructing the agent to read the relevant client's CLAUDE.md before any account-specific MCP call, is a prompt-based enforcement mechanism. The open question the poster raises, whether this kind of account fencing should instead be enforced deterministically through hooks or permission rules rather than relying on the model to follow instructions, points to a real gap in how Claude Code handles multi-tenant data isolation. Prompt-based rules are inherently probabilistic and prone to failure under context pressure or model inattention, which is a meaningful risk when the goal is preventing cross-client data leakage, an outcome that could have real business and compliance consequences for an agency handling multiple clients' Google Analytics, Google Ads, and Search Console data through MCP integrations.
The second half of the post addresses a subagent delegation strategy that has become increasingly common in the Claude Code community: reserving the most capable model for architectural thinking and judgment calls while offloading "grunt work" to Sonnet-based subagents. The poster's specific implementation uses two subagents, an implementer for well-specified coding and document tasks and a data-runner dedicated to MCP pulls, with the explicit goal of keeping raw JSON output from tools like GA4 or Google Ads out of the main context window. This reflects a broader pattern in agentic AI development where context window management is treated as a first-class engineering concern rather than an afterthought. Bulky, mechanical, or output-noisy work gets pushed to isolated subagent contexts that return condensed summaries, while the main thread preserves specs, architecture, and higher-order reasoning. The tradeoff the poster identifies, that subagents "start blind" and therefore aren't suited for small in-context edits requiring shared understanding, is an important nuance: delegation isn't free, and matching task type to agent capability requires judgment about how much shared context a task actually needs.
Collectively, this thread illustrates how sophisticated non-engineering users, in this case a solo marketing operator rather than a professional software team, are pushing Claude Code into enterprise-like multi-tenant use cases and running into infrastructure questions that Anthropic's official documentation doesn't fully address. The tension between prompt-based governance and deterministic enforcement mechanisms (hooks, permission rules) is emblematic of a broader challenge in the agentic AI space: as these tools move from single-user coding assistants toward systems managing sensitive, siloed data across many clients or business units, the demand for hard guarantees rather than soft instructions grows. This mirrors similar conversations happening around Claude's use in regulated industries and multi-client consulting scenarios, where the community is essentially crowdsourcing best practices that vendors haven't yet formalized, and where subagent orchestration patterns like the one described here are increasingly treated as a mature technique for managing cost, context limits, and task specialization simultaneously.
Read original article →