How to use the Create Implementation Plan agent skill (create-plan)

What is an agent skill, and what does create-plan do?

An agent skill is a folder containing a SKILL.md file with instructions that an AI agent loads when a request matches the skill's description. The create-plan skill (licensed Apache-2.0) tells the agent to plan rather than build.

When you ask for a plan, roadmap, or strategy for a complex change, the agent first researches the codebase using search, semantic search, and read tools, then writes a Markdown plan file in a plans/ directory. The plan contains an objective, a checkbox task list, verification criteria, risks with mitigations, and alternative approaches.

• Plan file name pattern: plans/{YYYY-MM-DD}-{task-name}-v{N}.md
• Implementation tasks use checkbox format (- [ ]) so you can track progress
• Source references use filepath:line format
• The skill forbids editing code, writing code snippets in the plan, and running tests or builds

Install the create-plan skill

Download the ZIP of the create-plan skill from this page. Then install it using one of these two options.

Option 1 is a personal install. Unzip the file into ~/.claude/skills/ so the skill is available across your projects. Option 2 is a project install. Unzip the file into the .claude/skills/ folder of a specific project so only that project uses it.

Option 3 is the Claude apps. Upload the ZIP under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Project: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills

Trigger the skill with natural requests

You do not need a special command. The skill is chosen when your request matches its description. Ask clearly for a plan, roadmap, or strategy, and name the change you want planned.

The skill's description mentions three situations: when you need thorough analysis before implementation, when you are breaking a complex feature into actionable steps, and when you explicitly ask for a plan, roadmap, or strategy.

• "Create an implementation plan for adding email notifications to the order service."
• "Give me a roadmap for splitting the monolith's billing module into its own service."
• "What strategy should we use to migrate our auth to a new token format? I want a plan before any code is written."

Requirements and what the plan needs

The skill needs an agent runtime that supports SKILL.md skills and provides search, read, and file-writing tools. It also needs a Bash shell, because the validation scripts validate-plan.sh and validate-all-plans.sh run in Bash.

After the plan is written, the agent runs ./.forge/skills/create-plan/validate-plan.sh on it, checks naming and structure, and fixes reported errors until the file passes. The validator checks, among other things, the filename, that the file sits in plans/, and that required sections are present.

• Agent runtime with SKILL.md skill support
• Search, read, and file-writing tools
• Bash shell for the validation scripts

Safety notes: scripts and network

The skill runs scripts: the bundled validation scripts are executed by the agent through Bash. Review the scripts before you install if your environment requires it.

The skill does not use the network, according to its listing. Its job is to read your files and write a plan file. It is designed not to change your code, so your source files should not be modified by this skill.

• Runs scripts: yes (validate-plan.sh, validate-all-plans.sh)
• Uses network: no
• Writes plan files under plans/; it is not meant to edit source code

Tips for better plans

Give the agent a concrete goal and scope. The skill says to make reasonable assumptions when requirements are ambiguous and to document those assumptions in the plan, so state anything you already know to avoid unwanted assumptions.

Review the plan before implementing. Because the plan is a file, you can edit it, check off tasks, and hand it to an implementation agent. The skill itself suggests switching to an implementation agent when you want code written.

• State the goal, scope, and any constraints in your request
• Ask for alternatives if the trade-off matters to you
• Read the risks and verification criteria before approving the plan
• Use a separate implementation agent for the actual coding