Menu
Allows an agent to hire a tmux minion
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Transform an application's logging/output to Brandon's colorized logging style — a reset-safe ANSI color layer with semantic level functions (info=yellow, warning=bold-yellow, success=green, error=red, verbose-trace=magenta) where color alone carries the level and there are no text labels. Use when the user says "apply my logging style", "colorize the logging", "make the logging match my style", "convert print/console.log statements to my colored logging", "add my logging library", "standardize this app's log output", or wants a project's ad-hoc prints and stdlib logging replaced with this specific style. Works in any language; Python is the reference.
anthropic's guidance for prompting llms
Address open PR review findings with judgment — read every finding, decide whether the feedback is right, fix or push back, resolve, and re-review by pushing. Repeat until clean. Use when the user says "address the PR review", "handle the review threads", "go through the review comments", or asks to respond to PR feedback on a specific PR or the current branch's PR.
Install the agent code-review GitHub Action into the current repository — writes the .github/workflows/code-review.yml workflow and sets the DEEPSEEK_API_KEY repo secret from the macOS keychain. Use when the user says "set up agent code review in this repo", "install the code review action", "add AI code review to this repo", "set up the AI code reviewer here", or wants automated PR review wired into the repo they're currently in.
Writes a message to a future session's agent. Send them your fondest memories of what you worked on, what went well, what didn't go well, and any useful tips they might need. You always run this when you finish a unit of work (closed a PR, completed the handed task, etc) or cross 300k–350k tokens of context. ALWAYS.
This skill should be used when the user asks to "run a slash command in another Claude session", "trigger /clear (or /compact, /model, /context, /rewind, …) in a tmux pane", "drive built-in commands in a running agent session over tmux", "send a slash command to the Claude/codex/opencode instance in session X", or otherwise wants to execute a harness command that the agent running in that session cannot invoke itself. Distinct from tmux-talk: this sends the raw command with no message envelope.
Based on SOC occupation classification
| name | hire-a-minion |
| description | Allows an agent to hire a tmux minion |
Hire a minion: a second claude instance, running on its own git worktree, in a new tmux window in your session. You stay the controller — you plan, groom, and review; the minion does the implementation work in parallel.
MINION=~/.claude/skills/hire-a-minion/bin/hire-a-minion
# Fresh minion: auto-creates a worktree, runs opus
ADDR=$("$MINION")
# Resume an existing session — its worktree name is REQUIRED
ADDR=$("$MINION" --worktree feature-auth --resume <session-id>)
# Pick the model, override the system prompt
ADDR=$("$MINION" --model sonnet --append-system-prompt "Be terse. Ask before large refactors.")
The script prints one line on stdout: the minion's tmux address
(session:window.pane). Capture it. Everything you do to the minion afterward
goes through that address. A worktree is always used — there is no way to run a
minion without one. It fails loudly if you are not inside tmux.
Flags: --worktree [name] (omit name to auto-create), --model sonnet|opus
(default opus), --resume <session-id> (needs --worktree <name>;
--continue is intentionally unsupported), --system-prompt,
--append-system-prompt.
Use /tmux-talk for every message to and read from the minion — feed it the address the spawn printed:
TALK=~/.claude/skills/tmux-talk/bin/tmux-talk
$TALK wait "$ADDR" 60 # let it boot before the first message
$TALK send "$ADDR" "<your brief>" # brief it
$TALK read-screen "$ADDR" 300 # see what it's doing
Talking and commanding are different seams. /tmux-talk carries conversational
messages — wrapped in a reply envelope — so use it to brief and read. /tmux-command
injects the minion's built-in slash commands (/clear, /compact, /model, …): the
minion's own agent cannot trigger these itself, and they must arrive unwrapped to be
recognized. Never send a slash command through /tmux-talk — its envelope puts the command
on a non-first line, so the CLI reads it as chat, not a command.
CMD=~/.claude/skills/tmux-command/bin/tmux-command
$CMD send "$ADDR" "/compact" # free its context without losing the thread
$CMD send "$ADDR" "/model opus" # switch its model mid-task
$CMD read-screen "$ADDR" 40 # confirm the command took
Stop a minion that's heading the wrong way without clearing its context:
$CMD keys "$ADDR" Escape
The minion exists to help you accomplish more work. It runs on a worktree, in parallel with you. While it implements, you ideate, plan the next unit, groom the backlog, and review the last result. Your responsibilities:
wait then read-screen to
see claude booted in the right worktree and isn't stuck. Don't assume — look./address-pr-reviews itself. You still own which work it
pulls and the order things merge./message-in-a-bottle) or
cleared, clear it for it: $CMD send "$ADDR" "/clear". Re-brief after
clearing — it loses everything.Escape to interrupt, read-screen to diagnose,
then either send a correction, or /clear and re-brief, or kill and respawn
with --resume <session-id> plus the worktree name to pick up where it left
off.$TALK idle "$ADDR" / $TALK wait
first — keystrokes sent mid-generation queue up and get misread.tmux kill-window -t "$ADDR". Don't leak
worktrees or orphan windows.Hire more than one when you have independent, non-overlapping work. Each gets its own worktree and window, so they don't collide. The limit is your attention: you must still brief, track, and review every one. Two minions you review well beat five you don't.