cplan

name	cplan
description	Call Claude (your existing Claude Code login — whichever account/host) to produce a plan, with optional custom system prompt. Default system prompt is the /plan skill (research-first, design-primer-aware, Qwen-ready format, recommends an executor model). Uses the local `claude` CLI so no API key juggling. Use when user says "/cplan", "claude plan", "cplan with system prompt", or wants Claude-as-planner without setting up keys.
category	execution
runtimes	["claude"]
pii_safe	true

/cplan — Claude planner via local claude CLI

A thin wrapper around the claude CLI on the current machine. Uses whichever Claude Code session you're already logged into (<MAC_HOST>, <WORKSTATION_HOST>, or wherever) — no API key file, no header juggling, just shells out to claude --print --model <model> --permission-mode plan.

By default the system prompt is the /plan skill content, so every cplan invocation gets:

• Step 0: Memory pass (/load + mcporter + prior plans + project tags)
• Step 1: Online research (versions, references, API docs, gotchas)
• Step 2: Clarify (only if genuinely ambiguous)
• Step 3: Write the plan in Qwen-ready format
• Step 3.5: Recommend the local executor model (qwen, coder-next, qwen-next, etc.)
• Step 4: Save plan to memory
• Step 5: Hand off

Tool

The cplan shell command exists on all 3 machines (<WORKSTATION_HOST>, <INFERENCE_HOST>, max). It locates and invokes claude from PATH or common install dirs. Output goes to stdout (or -o <file> to save).

Usage

# Default — use /plan skill as system prompt
cplan "build a settings page for <PROJECT_NAME> with dark mode toggle"

# Save the plan to a file
cplan -o ./plans/<PROJECT_NAME>-settings.md "build a settings page..."

# Custom system prompt (inline)
cplan -s "You are a senior security engineer. Produce a threat model." "review the auth flow in <PROJECT_NAME>"

# Custom system prompt from a file
cplan -s ./prompts/threat-model.md "review the auth flow in <PROJECT_NAME>"

# Override model (default is opus)
cplan -m sonnet "quick refactor plan for the discord realtime daemon"
cplan -m opus "complex architectural decision for the <PRODUCT_NAME> ingestion pipeline"

# Plain chat (no system prompt at all)
cplan --plain "what's the current best way to do server-side streaming in Next.js 15?"

When to use vs alternatives

Skill	Provider	Account	When to use
/plan	model-agnostic	depends on opencode/PI selection	You're already in opencode/PI and your model is set; this is the system prompt that drives whoever's selected
/cplan	Claude (CLI)	your current claude login	You want Claude specifically, from terminal, without worrying about which provider/billing
/oplan	main Claude (this session)	this session's account	Already in Claude Code; want Opus + full repo context inside the session
/pplan	local Ollama via PI	free / local	Sensitive plans, no cloud round-trip needed

Behavior when invoked

When /cplan is called:

1. Take the user's task as $ARGUMENTS.
2. If a custom system prompt was provided (text or file), pass it via -s.
3. If the user asked for a specific output file or a slug-style save, pass -o ./plans/<slug>.md (mkdir if needed).
4. Run via Bash:

   cplan [-s ...] [-m sonnet|opus] [-o ...] "<task>"
   

5. Surface the plan output to the user. If saved to a file, also print the path + the recommended executor command (lms-warm <model> "<system>" for warm-up + handoff).

Notes

• Uses local claude CLI — no API key. Whatever account claude is logged into is what gets billed.
• Default model is opus. Switch to sonnet for faster/cheaper plans.
• Output is plain text to stdout unless -o is passed.
• --permission-mode plan is set so claude doesn't try to write files — it just produces the plan text.
• Pair with /pdo to execute the plan locally on the recommended Qwen model.

Related

• /plan — the system-prompt content this skill uses by default
• /pdo — execute a saved plan with local Qwen
• /oplan — Opus plan inside Claude Code (current session, full context)
• /pplan — local-only plan via PI + Ollama
• /lms-warm — pre-load + prime the recommended executor model