Planning With Files Zh

When you need to architect enterprise Azure infrastructure from scratch rather than deploy individual apps, this skill walks you through the full planning process. It pulls Azure best practices and Well-Architected Framework guidance via MCP tools, generates detailed infrastructure plans as JSON, then outputs production-ready Bicep or Terraform code. Handles complex scenarios like hub-spoke networks, multi-region DR topologies, and compliance-driven designs that span multiple resource groups or subscriptions. The workflow enforces approval gates before any code generation, so you review the architecture plan before committing to infrastructure-as-code. Best for platform engineers who need systematic infrastructure planning rather than quick app deployments.

Forces Claude to actually think before coding, which frankly should be the default but isn't. Blocks all implementation until you've walked through context exploration, clarifying questions, approach options, and design approval. Generates proper spec docs with self-review checks for placeholders and contradictions. The hard gate against "this is too simple to design" saves you from those deceptively basic requests that turn into scope creep nightmares. Most useful when you're tired of Claude jumping straight into half-baked solutions that miss obvious requirements or architectural decisions.

Takes specs and generates step-by-step implementation plans with actual code snippets, exact file paths, and specific test commands. Maps out file structure first, then breaks everything into 2-5 minute tasks following TDD cycles. Saves plans as markdown with checkbox syntax in `docs/superpowers/plans/`. Runs self-review checks for placeholder content and type consistency. Works best when you have clear requirements but need to think through the implementation approach before coding. The bite-sized granularity keeps you from getting stuck on large, vague tasks.

Takes a written implementation plan and executes it step by step with proper checkpoints. Loads the plan file, reviews it critically for gaps or unclear instructions, then works through each task while marking progress and running verifications. Stops immediately when it hits blockers rather than guessing, which prevents the common problem of AI assistants making assumptions that derail implementations. Creates a TodoWrite file to track progress and integrates with git worktrees for isolated development. Really shines when you have a solid plan but want systematic execution with built-in quality gates.

Takes a fresh repo and does the boring first pass: reads the README, scans for setup scripts and documented commands, then categorizes what looks like inference versus training versus evaluation workflows. Outputs a structured plan with the safest reproduction path forward. Saves you from manually parsing through project files to figure out where to start, especially when you're dealing with academic ML repos that have scattered documentation. The classification tends to be conservative, which is good since it's meant to feed into an orchestrator that makes the actual execution decisions.

Planning-with-files implements a file-based task management system that helps users organize and track progress on complex multi-step projects by maintaining three markdown files (task_plan.md, findings.md, progress.md) that persist across sessions. It's designed for anyone tackling research tasks, multi-step projects, or work requiring numerous tool calls, solving the problem of losing context and progress when switching between sessions or managing intricate workflows. The skill automatically restores prior context and provides structured checkpoints for tracking phases, findings, and updates throughout a project's lifecycle.