Detailed Analysis
A Reddit user's detailed workflow post for Claude Code, built around what they call "claude-code opus 5," has surfaced a practical pattern gaining traction among developers working on multi-session coding projects: replacing formal "skills" or plugin systems with a structured hierarchy of markdown files. The setup consists of four files—claude.md (persistent instructions loaded every session), roadmap.md (the overarching project plan broken into phases with tasks assigned to agent, owner, or shared responsibility), todo.md (the active phase's checklist, updated at the end of each session), and handoff.md (a fully rewritten summary at the close of every session that primes the next one). The author reports that after burning through initial credits dealing with a disorganized codebase, this file-based memory system solved persistent context-tracking problems for both themselves and the model.
The post is notable less for the specific tool names—likely referring to Claude Code paired with an Opus-tier model, with "opus 5" being either informal shorthand or a forward-looking/aspirational reference given no such official release exists as of this writing—and more for what it reveals about how practitioners are working around the inherent context limitations of long-running agentic coding sessions. Claude Code, like other agentic coding tools, loses working memory when a session is cleared, and users have converged on markdown-based external memory as a lightweight, transparent, and version-controllable alternative to relying on the model's internal context window or proprietary "skill" abstractions. The technique effectively turns the filesystem itself into a persistent state machine: each session begins by reading handoff.md, executes against todo.md, and updates both before termination, creating a rolling audit trail that also serves as documentation.
The article's more pointed claim—that a third-party "superpower" skill wasted significant weekly credit allowance on an unproductive tangent, while an "ADHD" skill (evidently a widely circulated community prompt for improving focus and task adherence) worked better when pasted directly into claude.md rather than loaded as a discrete skill—speaks to a broader tension in the Claude ecosystem around Anthropic's skills and plugin architecture. Skills are designed to modularize and reuse specialized instructions or tool integrations across sessions, but this anecdote suggests that community-authored skills can introduce unpredictable, resource-intensive behavior, whereas directly embedding vetted instructions in the model's core context file offers more control and lower risk. The user's explicit endorsement of only using MCP (Model Context Protocol) integrations from primary infrastructure providers—Vercel, Supabase, GitHub—over generic community skills reflects a growing wariness in the developer community about extending agent capabilities through unaudited third-party extensions, especially as usage-based credit systems make wasted agent runs directly costly.
More broadly, this post fits into a pattern seen across the Claude Code and agentic-coding community: as these tools mature, users are developing increasingly sophisticated "operating procedures" around the models rather than waiting for the underlying systems to natively solve memory, planning, and hand-off problems. The emergence of grassroots conventions—phased roadmaps, standardized handoff documents, session-scoped todo lists—mirrors software engineering discipline being reapplied to human-AI collaboration itself. It also underscores a recurring theme in AI-assisted development: raw model capability is necessary but insufficient, and the surrounding scaffolding (context management, task decomposition, credit-conscious tool selection) often determines whether an agent is productive or wasteful. As Anthropic continues to iterate on Claude Code's skills and plugin ecosystem, feedback like this suggests real demand for more curated, reliable extensions and better default context-management primitives, rather than an ever-growing marketplace of community skills whose quality and resource efficiency vary widely.
Read original article →