Preface¶

When modifying a repository with an Agent, a common scenario is that the model writes files while analyzing: the diff has already landed before the solution is aligned. Rollbacks, requirement comparisons, and supplementary testing become expensive. If you want the Agent to first break down the task clearly, articulate the risks explicitly, and modify the code only after you nod, you need a repeatable “planning gate.” This article introduces the SkillHub community skill Plan Mode (@user_5607203c/plan-mode), triggered by /plan, which constrains the session to a “plan first, execute later” read-only analysis workflow.

What is this¶

Plan Mode is a community Skill published by user_5607203c on SkillHub (Category ai-agent, Version 1.0.0, Tag requires_api_key=false). It solves the following problem: before the user explicitly approves, the Agent should not perform any operation that modifies the repository or environment; it must first output a structured plan and wait for user confirmation before proceeding to implementation.

Compared to “letting the model modify code directly,” this Skill embeds PLAN MODE into the skill instructions: upon triggering, it defaults to read-only analysis, and modification operations require approval.

Core Functions¶

Based on the SkillHub page and API materials, the verified capabilities are as follows:

1. Trigger Method: When a user inputs /plan, the PLAN MODE defined by this Skill is enabled.
2. Plan First, Execute Later: In PLAN MODE, it operates as a read-only analysis mode; planning output must be completed before considering execution.
3. Plan Structure: Before approval, a complete plan must be output, containing:
   • Task Understanding
   • Context Analysis
   • Implementation Plan
   • Risk Points
   • Verification Scheme
4. Approval Gate: Any modification operation (writing files, changing configuration, running commands that change state, etc., subject to skill constraints) can only be executed after waiting for the user’s explicit approval.

Installation and Activation¶

SkillHub provides two common installation paths: hand the installation prompt to an AI client that supports Skills, or use the skillhub CLI to install it to the current Agent’s skills directory.

Method 1: Installation via AI (Official Page Prompt)¶

Send the following intent to the current Agent (installation steps are subject to SkillHub documentation):

Please install `@user_5607203c/plan-mode` based on https://skillhub.cn/install/skillhub.md.

Method 2: skillhub CLI¶

First, confirm if the CLI is available:

command -v skillhub && skillhub --version

If not installed, follow the SkillHub installation documentation to perform a full installation (example from official skillhub.md):

curl -fsSL https://skillhub-1388575217.cos.ap-guangzhou.myqcloud.com/install/install.sh | bash

When installing the skill, you must use --dir to point to the current Agent’s skills directory; otherwise, it defaults to ./skills/, which might not be loaded by the client. Common path examples (excerpted from SkillHub installation docs):

Install the coordinate @user_5607203c/plan-mode to the corresponding directory. The command format is as follows (replace <skills directory> with the path from the table above):

skillhub install @user_5607203c/plan-mode --dir <skills directory>

After installation, refresh or restart according to your client’s conventions to load the Skill. This Skill does not require an API Key.

Typical Usage¶

Below is a reproducible conversation flow corresponding to the trigger and gate descriptions in the official introduction.

1. In an Agent that has loaded this Skill, input /plan first for a specific task and explain the goal (e.g., refactoring a module, adding an interface, fixing a bug).
2. Wait for the Agent to output a complete plan under PLAN MODE and verify if the five parts are complete: Task Understanding, Context Analysis, Implementation Plan, Risk Points, Verification Scheme.
3. If the plan is incorrect or the scope is too large, add constraints within the same mode and request a plan revision; you should still not see the repository modified before approval.
4. Approve with explicit wording (e.g., “Execute according to this plan”), then let the Agent enter the implementation phase.

Example (User input, for trigger demonstration only):

/plan
Please add a health check interface to the current project. Do not change the code first; provide a complete plan first.

Before approval, the Agent should remain in analysis and planning; modification operations are only allowed after approval.

Use Cases and Notes¶

Who is it for

• Developers who need to align scope, steps, and verification methods before modifying code.
• Teams undertaking high-risk changes (database migrations, permissions, release scripts) who want to go through risk and verification checklists first.
• Reading scenarios where you want to reduce “Agent modifying files arbitrarily” and facilitate pre-emptive Code Review.

Limitations and Notes

• This Skill constrains the Agent’s behavior under PLAN MODE; if the user sends ambiguous instructions after approval, the implementation phase may still deviate from the plan. It is advisable to clarify boundaries during approval.
• Read-only analysis depends on the client and model’s adherence to the Skill; for critical environments, it is still recommended to combine with branch protection and CI.
• SkillHub Statistics (based on scraped data): Rating 100000, Downloads 959; Community source, maintenance rhythm depends on the publisher’s page.
• If coexisting with IDE built-in “Plan” modes, refer to the actually loaded Skill and client rules to avoid mixing different entry points.

Conclusion¶

Plan Mode binds /plan with “Complete Plan + Explicit Approval,” making it suitable for Agent collaboration where you “think clearly before acting.” To install or view instructions, you can open the SkillHub page:

• https://www.skillhub.cn/skills/user_5607203c/plan-mode
• https://api.skillhub.cn/user_5607203c/plan-mode