/wtf — "What's The Flow?" (your what-do-I-do-now helper)

WTF = "What's The Flow?" — the one thing to type when you're stuck on which command to run next. Three modes, picked from what was typed:

Be decisive: recommend ONE thing (not a menu) — a bold mini-header, a one-line why, the exact thing to type, and a "want me to run it?" A menu is exactly what confuses someone who's stuck.

MODE 1 — RECOMMEND (the default)

Don't guess where they are — check, then map state → the single best next step. Run these (cheap, read-only):

pwd
git rev-parse --abbrev-ref HEAD 2>/dev/null     # current branch
git status --porcelain 2>/dev/null              # dirty? (any lines = uncommitted edits)
git log --oneline -1 2>/dev/null                # last commit
git status -sb 2>/dev/null | head -1            # ahead/behind origin

Also read the conversation so far — half the signal: have they just opened the tab? been editing? been debugging? winding down? Combine git state + what they've been doing to pick.

State → next step (check top to bottom; first match wins)

1. Fresh tab / not oriented / "where am I" → /start <topic> — lands you on the right branch, clean, with the resume note loaded. Always the first move in a new chat.
2. Edited + tree dirty + wants to save/back up → /commitpush — focused commit + push.
3. Winding down / about to /clear → /end — /wrapup then /handoff, back-to-back.
4. Just wants to share/format what they figured out → one of /prettymsg (read it now), /prettyhtml (keep it), /discordmsg / /discordembed (share it).
5. Behind origin / "get me current" → git pull --ff-only (if clean).

Adapt this list to whatever skills the user actually has installed (look in ~/.claude/skills/). If two could fit, prefer the one matching what they just said they want. If genuinely ambiguous, ask one short either/or — never dump the table.

How to answer (RECOMMEND mode)

Keep it to ~4 lines, tidy:

## 👉 Run this: /commitpush
You've got uncommitted edits and you're about to switch tasks.
It stages just those files, commits with a clean message, and pushes.
Want me to run it now?

Offer to run read-only/safe commands immediately; for anything destructive or that pushes to a shared place, recommend it and wait for an explicit "yes."

MODE 2 — EXPLAIN

They named a command ("/wtf explain handoff", "what does wrapup do?"). Give a plain, friendly, non-jargon explanation: what it does, when to use it, and the one gotcha. The source of truth for each skill is its own SKILL.md — read it (~/.claude/skills/<name>/SKILL.md, or the project's .claude/skills/<name>/SKILL.md) if unsure; don't invent behavior.

MODE 3 — MAKE a new skill

They said "make a skill that <does X>" / "write a skill for <X>". Scaffold it so it shows up as /<name> and behaves like the others.

1. Clarify only if needed (max 2 short questions): the name (/<name> they'll type) and personal vs project if it's not obvious. Default to personal (~/.claude/skills/, available everywhere) unless it's clearly a shared/repo workflow skill (then the project's .claude/skills/). Infer from X; don't over-ask.
2. Write <dir>/<name>/SKILL.md with the standard frontmatter:

   ---
   name: <name>
   description: <one paragraph — what it does + when, INCLUDING the trigger phrases the user would
                say, because the description is what Claude matches on to fire the skill>
   user-invocable: true
   argument-hint: [optional — what args it takes]
   ---
   
   # /<name> — <one-line title>
   
   <clear, step-by-step instructions to the model on exactly what to do when invoked.>

   • Match the voice and shape of the existing skills (look at prettymsg / start as models).
   • The description is load-bearing: pack the natural trigger phrases into it.
3. Tell them it's live: "type /<name> — it'll show up next prompt," show the trigger phrases, and offer a dry-run.
4. Do NOT commit anything unless asked.

Always

• One recommendation, not a menu.
• Tight + skimmable — a bold mini-header, a couple emojis.
• Never commit/push unless explicitly asked (that's what /commitpush is for).