How to use the leader agent skill: turn ideas into agent briefs

What the leader skill makes the agent do

You describe a goal in one sentence. The skill first inspects your code repository and runs real commands to check the facts, then searches the web if it needs outside information. It asks you at most five focused questions, and then writes a brief of at most 4000 characters.

The brief is written to be pasted straight into goal mode of a separate agent. That agent works through the project without you present. The brief also plans how you will check the result afterwards, including anti-cheating checks, progress notes so work can resume after an interruption, and a rule for splitting a large job across several agents in parallel.

• Tasks that change code get a brief with verifiable completion conditions.
• Exploratory tasks such as research, choosing a tool, or finding an approach are routed to a different brief format automatically.
• The skill is not meant to replace your review of the result.

How to install the leader skill

Download the ZIP file from this page and unzip it. Then put the skill folder in one of two places, depending on how you use Claude.

For Claude Code, place it in your user folder at ~/.claude/skills/, or inside a single project at that project's .claude/skills/ folder. In the Claude apps, you can instead upload the ZIP under Settings, then Capabilities, then Skills.

• Personal use across all projects: ~/.claude/skills/
• One project only: the .claude/skills/ folder inside that project
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

How to trigger the skill with natural requests

You do not need a special command. Ask in plain language, and the skill activates on requests that match its description. The skill's description lists these example phrases, which are in Chinese in the source:

• 帮我给 agent 写个目标 (help me write a goal for an agent)
• 帮我详细拆一下这个目标 (help me break this goal down in detail)
• 写个任务书/brief 给 agent (write a task brief for an agent)
• 写个 goal 提示词 (write a goal prompt)
• 让 agent 自己跑这个项目 (let an agent run this project by itself)
• 把活分给几个 agent 并行 (split the work across several agents in parallel)

Example workflow from idea to finished result

First, ask for a brief using one of the phrases above and describe your idea in a sentence. Answer the questions the skill asks, which are limited to five and are chosen because they cannot be found out by research and would change the brief. If you do not answer one, the skill proceeds with a default and lists that default in the brief under a section where you can review it.

Second, open the agent that will do the work and type /goal followed by a space. Paste the whole brief that the skill produced. There is no separate input box for the command, and tools without goal mode can take the pasted brief directly.

Third, when the agent finishes, return and tell the skill. It then re-runs the verification commands in the brief and gives you a short plain-language report covering whether it passed, what was done, open issues, and next steps.

• Before pasting, skim the section that lists the decisions made on your behalf, and the two hard completion conditions at the end.
• Do not send a second /goal command for the same job; the skill is designed for one goal and one paste.

What the skill needs and its safety notes

The skill needs Claude Code or another agent that has a goal mode (the /goal command), or any agent you can paste the brief into. It also needs access to the target code repository so it can run real commands to check claims and verify results. The skill runs no scripts of its own, but it does use the network, for example to search for background information.

The skill itself warns that agents can misreport results or game tests. A common example is making a failing test pass by skipping it or weakening its assertions instead of fixing the code. The brief names these tricks and sets a baseline so they count as failures. You should still verify the results yourself before relying on them.

• Requires: Claude Code or another agent with goal mode, and access to the repository
• Runs scripts: no
• Uses network: yes
• Do not use the brief without human review of the results

Tips for getting good briefs

Give the skill a real repository to inspect. It checks whether documented commands exist, what the baseline numbers are, and where the documentation differs from reality, so a brief written without that check is weaker.

Keep one job per brief. If the brief would exceed the 4000-character limit, the job is too big and should be split into separate pieces, each given to an agent in its own paste.

Only ask for parallel agents when you are ready to answer the questions up front, because the skill asks for that decision in its first round of questions.

• Describe the outcome you want, not the method, and the skill will turn the outcome into numbers.
• If a task cannot be checked by a command, expect the skill to mark it as something that needs a person to check it.
• Read the progress file the agent writes if work is interrupted, rather than starting over.