Using chief-of-staff to Coordinate Sub-Agents and Drive Long-Term Goals Forward
機械翻訳 / Machine-translated
Original article: Using chief-of-staff to Coordinate Sub-Agents and Drive Long-Term Goals Forward (cccskills)
Skill page: chief-of-staff
Key takeaways from this article
With chief-of-staff installed, agents in Claude Code or Codex become coordinators for long-term goals.
Work is delegated to sub-agents, while the coordinator itself continues interacting with you.
This article explains how to keep records using the Wiki and how to think about improving your working environment.
The latter half also covers installation steps and example prompts.
When you invoke chief-of-staff, the agent acts as the person responsible for your goal.
It's a coordinator that doesn't do the hands-on work itself — it delegates tasks to sub-agents.
Here's an example request:
/chief-of-staff I want to work toward migrating our internal admin panel to Next.js across multiple sessions.
Please take on the coordinator role.
The first thing that comes up is a discussion about where to store the Wiki.
The Wiki is a folder of interlinked Markdown pages that serves as persistent memory across sessions.
For a goal that doesn't yet have a Wiki, the coordinator will first discuss the location with you before writing anything.
Once the location is decided, actual work is delegated to background sub-agents.
Because the conversation with the coordinator stays open, you can discuss changing priorities or ask questions at any point during the migration.
Instructions passed to sub-agents are kept concise.
Information is shared by pointing to research notes, past commits, and Wiki page locations — not by copying content inline.
A defining characteristic of this skill is that before starting any work, the agent always first considers how to improve the environment it operates in.
Alongside the migration work, you'll receive suggestions for adding lint rules to prevent mistakes and making development server logs readable.
chief-of-staff is a skill that gives an agent a complete approach for following a long-term goal across sessions.
It covers five areas:
The coordinator maintains two perspectives in parallel.
One is a tactical perspective: "How do we handle the problem right in front of us?"
The other is a strategic perspective: "How do we change the environment so the next problem goes more smoothly?"
All work is performed by sub-agents, keeping the coordinator's context window (the amount of conversation it can hold at once) free.
Because agents run in the background, dialogue with you is never interrupted.
To avoid repeating explanations, existing materials are shared simply by referencing their location.
Knowledge that exists only within a conversation is lost when the conversation is compressed.
To address this, the coordinator writes decisions, recurring issues, and user preferences to the Wiki — one topic per page.
The index.md holds a list of all pages along with a one-line summary of each.
Source materials such as issues, pull requests, and sub-agent reports are cited as references rather than copied verbatim.
There are three operations for keeping the Wiki useful:
| Operation | When | What It Does |
|---|---|---|
| Ingest | When something long-lasting is learned | Updates all relevant pages and index.md |
| Query | Before making a decision | Reads index.md and referenced pages; turns answers worth keeping into their own page |
| Lint | Periodically | Looks for contradictions, orphaned pages, broken links, and outdated content |
A rule of the Wiki is to write as a history, not just a current state.
Entries include dates, such as "Decided on 2026-10-13 to do X."
For entries about code, the commit where that information was accurate is noted alongside the repository name.
Old entries are never overwritten when circumstances change.
A new dated entry is added, and the old one is marked as superseded.
The coordinator places emphasis on building an environment where agents can work correctly without undue friction.
Seeking out opportunities for improvement with every message you send is also part of its expected role.
Areas for improvement include:
It also enforces a rule of "no workarounds."
Ad hoc hacks are not permitted, and deviations from standards are to be corrected before feature development proceeds.
If the runtime environment supports it, the coordinator will suggest recurring schedules that help achieve the goal.
Wiki linting is one of the operations it may propose as a scheduled task.
Run the following command in your terminal.
You can choose which agent to install into — such as Claude Code or Codex — during the process.
npx skills add https://github.com/mattpocock/skills --skill chief-of-staff
If you'd like to have the agent handle the installation for you, send the following message as-is.
It can take care of everything from reviewing the contents to placing the files.
Please install the following Agent Skill.
https://github.com/mattpocock/skills/tree/main/skills/in-progress/chief-of-staff
First, check the contents of SKILL.md and any accompanying files to confirm there is no suspicious behavior (such as sending data externally or reading credentials), then place it in ~/.claude/skills/chief-of-staff/ for Claude Code, or ~/.agents/skills/chief-of-staff/ for Codex.
Choose the install location based on whether you want to use it across all projects or only in a specific project.
| Agent | For all projects | For a specific project only |
|---|---|---|
| Claude Code | ~/.claude/skills/ | .claude/skills/ in the project |
| Codex | ~/.agents/skills/ | .agents/skills/ in the project |
This skill is configured so the agent will not invoke it on its own initiative.
To use it, invoke it by name — for example, with /chief-of-staff in Claude Code.
Including the long-term goal and current situation in your request helps the conversation move quickly.
Here are three representative examples:
/chief-of-staff I want to build out tests for our service and reduce the amount of manual verification needed before releases.
Please organize who handles what for this goal and get it moving.
If there's no Wiki yet, the conversation starts with deciding where to put it.
After that, research and test creation are delegated to sub-agents, while the coordinator handles progress updates and discussions about the next decisions.
While the work is in progress, please also look into improving our standards and environment to make future work easier.
Separate from the ongoing work, suggestions will come back.
These might include adding lint rules to catch errors, improving CODING_STANDARDS.md, and making logs or test environment databases accessible.
Are there any recurring tasks that would be worth scheduling for this goal?
If the runtime environment supports it, recurring schedules like Wiki linting will be proposed.
In environments that don't support it, this suggestion won't appear.
Before using this skill, check your runtime environment and how permissions are handled.
The same mattpocock/skills repository contains skills that pair well with long-term work before and after using chief-of-staff.
claude-handoff is a skill that summarizes an ongoing conversation for handoff and passes it to a new background agent.
When a conversation grows long, it organizes the working approach and reference materials so the work can continue seamlessly.
code-review is a skill that reviews changes based on a specified commit or branch.
Because it reports separately on conformance to development standards and on whether the specification has been met, it's well suited to verifying that the standards established through chief-of-staff are being followed.
ask-matt is an advisor that recommends relevant skills for your situation and the order to use them in.
You can consult it whenever you're unsure which skill to use next.
No, it won't run on its own.
Because the skill is configured so the agent won't invoke it on its own initiative, you need to call it by name.
Important history is preserved in the Wiki.
Knowledge that exists only in the conversation is lost when compressed, so the mechanism is to write decisions and discoveries to the Wiki with dates and carry them over to the next session.
The skill's scope goes only as far as suggesting a recurring schedule when the runtime environment supports it.
Whether to actually set it up is for you to decide after seeing the suggestion.
chief-of-staff is a skill that lets an agent take on the coordinator role for long-term goals.
Work is delegated to sub-agents, freeing you to stay in dialogue with the coordinator and focus on making decisions.
History is preserved in a dated Wiki, and suggestions for improving standards and the working environment come with each request.
For more details, see the cccskills page.
This article is a reprint of the cccskills piece "Using chief-of-staff to Coordinate Sub-Agents and Drive Long-Term Goals Forward."
Full details on the skill and installation instructions can be found on the cccskills chief-of-staff page.