coach

Locate the scripts folder: ${CLAUDE_PLUGIN_ROOT}/scripts if that variable is set, otherwise the newest ~/.claude/plugins/cache/*/coachline/*/scripts (respect CLAUDE_CONFIG_DIR if set). Pick an interpreter: python3, else python, else py -3. On Windows, python3 may be a Store stub that prints nothing; if so use python.

• /coach (default): run <py> <scripts>/coach.py --coach --copy. Takes up to ~20s. It prints my enhanced prompt as one plain block and copies it to my clipboard; tell me that it did. Drop --copy if I say not to touch the clipboard, and add --no-llm if I say quick or offline.
• /coach review: analyse my whole history with my own Claude subscription (no API key).
  1. Run <py> <scripts>/review.py with no flags. It sends nothing and prints a plan. Show me the plan verbatim.
  2. Ask me: "Send these N redacted prompts to Claude through your subscription?" Wait for my answer in the conversation. NEVER run --yes on your own initiative or because an earlier step suggested it.
  3. Only if I say yes: run <py> <scripts>/review.py --yes (allow several minutes, ~20s per call; forward --days N / --max N if I asked). Show the report verbatim.
  4. Offer to install a suggested skill only by name, one at a time: <py> <scripts>/review.py --install <slug>. Never install without my saying which.
• /coach watch: the panel (one column, this thread only) must run in its own split pane, so you cannot run it inside this session. Tell me the short permanent command to paste in a second tab or pane: <py> ~/.claude/coach/panel.py (use the real path; run <py> <scripts>/launch.py --command to print it). ↑/↓ choose a prompt, Enter opens it full width, c copies its enhanced prompt, e analyses it, i shows patterns from your past chats, l opens the Library of saved prompts (c copies one, x forgets it), s installs a drafted skill, x hides a web suggestion, a marks one you use, r looks on the web again, m turns the mouse on or off, q quits.
• /coach panel-ai on / off: run <py> <scripts>/setup.py --panel-ai on (or off). It is on by default. While on, the panel itself sends each prompt of the current thread (redacted, with the thread's earlier prompts as context) and every ~2 days up to 90 days of redacted history to Claude in the background, with the names of installed skills and marketplace plugins as context, and, once per new task, a web research call (see /coach discover). Before running on, tell me that in one sentence and wait for my yes. Never enable it on your own initiative.
• /coach discover: research the web now for my last prompt: run <py> <scripts>/discover.py --last (allow up to ~4 minutes). Before running, tell me in one sentence that it sends a generic topic, my stack's framework names and versions, and my installed skill names to a web search through my Claude subscription, and wait for my yes. Show its output verbatim; it only lists suggestions that passed verification, and never install any of them yourself.
• /coach research on / off (old name: /coach discover on / off): run <py> <scripts>/setup.py --research on (or off). On by default while analysis is on: each new task gets one background web research call. Ask me before turning it on.
• /coach model fast|research haiku|sonnet|opus: run <py> <scripts>/setup.py --model <fast|research> <model>. Just do it; it changes which model my subscription uses for that pass.
• /coach auto-open on / off: run <py> <scripts>/setup.py --auto-open on (or off). While on, the panel opens by itself at session start: inside Claude Code on macOS and Linux (on by default), a Windows Terminal split on Windows. On Windows, ask me before turning it on; elsewhere just do it (it only shows the panel, it sends nothing).
• /coach hold on / off: run <py> <scripts>/setup.py --hold on (or off). Before on, tell me in one sentence that a clearly vague prompt will stop at Enter until I press Enter again, and wait for my yes.
• /coach level beginner|pro|auto: run <py> <scripts>/setup.py --level <value>. Just do it; it only changes how hints are worded.
• /coach hints on / off: run <py> <scripts>/setup.py --hints on (or off). Just do it; the hint is worked out locally and sends nothing.
• /coach motion on / off: run <py> <scripts>/setup.py --motion on (or off). Just do it; it only changes how the panel draws.
• /coach setup: run <py> <scripts>/setup.py with no options. It only shows the settings and the command that opens the panel.
• /coach doctor: run <py> <scripts>/coach.py --doctor.
• /coach forget library: deletes my saved prompts (the Library) and the ones waiting to be saved. Tell me that in one sentence and wait for my yes, then run <py> <scripts>/coach.py --forget library.
• /coach forget outcomes: deletes coachline's local record of interrupted and corrected prompts. Just do it: run <py> <scripts>/coach.py --forget outcomes.
• /coach forget all: deletes everything coachline keeps in ~/.claude/coach (analyses, caches, the history review, drafted skills) except my settings and llm-off.txt. Tell me that in one sentence and wait for my yes, then run <py> <scripts>/coach.py --forget all.

Show script output verbatim and add nothing: no rewriting of your own, no praise, no extra advice.