Claude Code skill
A way to hand Claude a plan that touches several codebases and have it run as a supervised project. Claude does the coding. You approve the plan, watch the board and sign off before anything reaches production code.
Think of it as a project manager and a small crew. The project manager (Mission Control) interviews you, writes a plan, and waits for you to approve it. Then it assigns one worker per codebase, tracks their progress on a live board, and routes any new to-dos you add to the right worker. When the work is done, an independent checker looks for clashes with what other people have shipped in the meantime. Only after you approve that check does the work go into the main codebase. The project manager itself never writes code.
Answer the intake questions, approve or send back plan sections, add to-dos, answer the occasional blocker, and give the final go-ahead.
Runs the interview, drafts the plan, holds the approval gates, launches and supervises the workers, merges finished work when your policy allows, and keeps the board current. Never writes feature code.
Do the actual coding. Each works only in its own repo, in a separate copy of the code, on its own tasks. Each submits finished work as pull requests and reports back in a fixed format.
Runs at the end. Compares the finished work against what has changed in the main codebase since the mission started and flags conflicts in code, shared contracts, config and docs. Cannot change anything.
The order below is fixed. Mission Control will not skip ahead, even if asked to "just start".
A short interview, one question at a time. Which repos will change and which are read-only context (both as tick-lists you can also answer by typing; the repo you are in and its dependencies come pre-selected), is there a written plan, what does "done" look like, who may merge what, and what is off limits. Secrets, production config and deploys are always off limits.
Mission Control reads the repos and splits the plan into small sections: the scope, a task list for each codebase, and the order things will be integrated. No code is touched.
The plan appears in a file called REVIEW.md. For each section you tick approve or changes and leave comments. Say "reviewed" in chat and Mission Control revises what you sent back. This repeats until every section is approved.
Mission Control creates a working area for each repo and starts one repo agent per codebase, all at once. Each agent gets only its own tasks and the approved scope.
Agents work through their tasks and open pull requests. The live board (MISSION.md) shows every task, its status and any blocker that needs you. You can add to-dos in TODO.md at any time and they are routed to the right agent. A task only counts as done when there is a merged pull request to show for it.
Full test suites run on the combined work. If you named a test environment, the work is deployed there and smoke-tested. A fresh verification agent with no prior context checks each "done" outcome and demands evidence. Failures become new tasks.
The read-only checkers look for clashes with anything merged to main since the mission began, drifted contracts between repos, outdated docs, and migration or config collisions. Findings land in REVIEW.md with a proposed fix each. You approve every section before anything moves to main.
Work is merged (or handed to you as a list of pull requests, depending on the policy you chose). You get a short summary: done criteria, findings and fixes, links, and anything still open.
Everything lives in plain text inside a folder called ~/.claude/missions/<mission-name>/ on your machine, outside the codebases themselves, so nothing about the mission ever gets committed. You can open it in any editor.
REVIEW.mdThe plan, section by sectionTick approve or changes, add comments, save. Later the pre-prod findings appear here too.MISSION.mdThe live boardPhase, progress, what needs you, one table per repo, recent activity. Mission Control rewrites it; you just read it.TODO.mdYour inbox to the crewAdd a line under the right repo. It becomes a task for that repo's agent and gets marked as routed.Two of the intake answers shape how much runs on its own.
This bundle contains both skills — mission-control-code (runs the mission) and mission-control-update (the refresh pass) — plus the two subagents they launch and an INSTALL.md. It works on any Claude Code install; no account or org access needed.
⇩ Download mission-control-skill.zip ~16 KB · both skills + INSTALL.md
Unzip it, then from the unzipped folder:
Install (once):
mkdir -p ~/.claude/skills ~/.claude/agents
cp -R mission-control-code mission-control-update ~/.claude/skills/
cp mission-control-code/agents/*.md ~/.claude/agents/
Restart claude so it picks up the new skills and agents. You also need git and the GitHub CLI logged in (gh auth status). Full notes are in INSTALL.md.
git and the GitHub CLI — run gh auth login once and choose GitHub.com so it can open pull requests for you.mission-control-skill.zip. You get a folder named mission-control-skill.cmd in the address bar.~/.claude/skills and the two helper agents into ~/.claude/agents.mission-control-code and mission-control-update folders into C:\Users\<you>\.claude\skills\, and the two files in mission-control-code\agents\ into C:\Users\<you>\.claude\agents\./mission-control-code. If it starts asking the intake questions, you're set.Open Claude Code in a folder that holds your repos (or an empty folder; it will fetch them), then type /mission-control-code or say "start a mission".
Any time you want a refresh, type /mission-control-update. It verifies what the agents reported, sweeps pull requests, routes new to-dos, merges what your policy allows, and gives you a three-line status. For a long mission, ask for it on a timer, for example every 20 minutes.