Skip to content
Hyprlayer

/create_plan

Create detailed implementation plans through an interactive, iterative process. The agent researches the codebase, works with you to understand requirements, and produces a phased technical specification.

Usage

Section titled “Usage”
/create_plan

Or with a ticket file:

/create_plan thoughts/shared/tickets/eng_1234.md

Process

Section titled “Process”
  1. Gather requirements — The agent asks for the task description, constraints, and context
  2. Research — Explores the codebase to understand relevant components
  3. Draft plan — Creates a phased plan with specific changes and success criteria
  4. Iterate — Works with you to refine the plan until it’s ready

Plan Structure

Section titled “Plan Structure”

Plans are saved to thoughts/shared/plans/ with the format YYYY-MM-DD-description.md. Each plan includes:

  • Problem statement and context
  • Phased implementation steps
  • Specific file changes per phase
  • Success criteria for each phase
  • Checkboxes for tracking completion

Key Behavior

Section titled “Key Behavior”
  • Uses Opus — Deep reasoning model for thorough analysis
  • Interactive — The agent asks questions and iterates on the plan with you
  • Skeptical — The agent challenges assumptions and identifies potential issues
  • Codebase-aware — Researches existing code before proposing changes

What This Adds

Section titled “What This Adds”

Claude Code has a built-in Plan mode, and GitHub Copilot has a similar planning flow. These are useful for quick, session-scoped planning — but they have real limitations for non-trivial work:

  • No persistent, shared artifact. Plan mode produces a plan in the conversation. You can press Ctrl+G to open it in an editor and save it manually, but there’s no automatic persistence, no shared location a teammate can review, and no artifact another session can pick up. /create_plan saves plans to thoughts/shared/plans/ as versioned markdown files — automatically persisted, team-accessible, and available to future sessions.

  • No phase structure or success criteria. Plan mode gives you a list of steps. /create_plan produces phased implementation plans with specific file changes per phase, success criteria for each phase, and checkboxes that /implement_plan marks off as work is completed. This creates an auditable trail from plan to execution.

  • No skepticism. Plan mode is accommodating — it takes your approach at face value and plans around it. /create_plan instructs the agent to be skeptical: challenge assumptions, identify potential issues, and push back on your approach when warranted. This catches problems before code is written, not after.

  • Shallow research. Plan mode spawns a single generic sub-agent for codebase reading. /create_plan spawns multiple specialized agents — codebase-locator, codebase-analyzer, codebase-pattern-finder, and thoughts-locator — in parallel, each focused on a different aspect of the research. It also searches your thoughts directory for prior plans, research, and historical context that Plan mode knows nothing about.

  • No iteration workflow. If you want to refine a plan in Plan mode, you continue the conversation and hope the model tracks the changes. With /create_plan, the plan is a file — you can review it offline, share it with your team, and refine it with /iterate_plan in a separate session that reads the artifact directly.

In short: Plan mode is for thinking out loud. /create_plan is for producing a spec that your team can review and your agents can execute.

Variants

Section titled “Variants”
  • /create_plan_nt — Skips thoughts directory integration
  • /create_plan_generic — Generic version for use outside Hyprlayer