--- name: triptych description: Orientation and setup-check — read where the user is in the work, ask about their goals, identify their position in the start → explore → work → check → wrap arc, surface a concrete recommended next step, and optionally check that Triptych is set up well for their use case. Use when the user types `/triptych`, `/?`, `/help`, asks "what should I do next?", "what commands do I have?", "where are we?", or seems lost mid-session. --- # /triptych — orientation `/triptych` doesn't automate decisions. It reads where you are, asks you what you're trying to do, and recommends a next move — leaving the choice with you. Use it whenever you're unsure what to do next, want a sanity check on setup, or just want to see the arc. The skill has three modes; default is the full interactive flow. - `/triptych` — interactive orientation (default). - `/triptych setup` — readiness check only (config + skills + MCPs for the user's use case). - `/triptych help` — static cheatsheet only, no interaction. ## Interactive flow (default) ### 1. Read state silently Before saying anything, gather context: ```python import sys; sys.path.insert(0, '.') from core.session import read_session from core.research import read_state, get_graph from displays import OUTPUT_DIR import json from pathlib import Path session = read_session() state = read_state() try: graph = get_graph() except Exception: graph = {'nodes': {}, 'edges': []} local_md = Path('CLAUDE.local.md').read_text() if Path('CLAUDE.local.md').exists() else '' output_files = [f.name for f in OUTPUT_DIR.iterdir() if f.is_file() and not f.name.startswith('.')] ``` Build a one-paragraph snapshot of where they are: - Goal (verbatim from session, or "no goal set yet") - Phase (exploration / formalization / no session) - Counts: claims emitted, established results, attempts, open threads - Loops running (verifier? watcher? dashboard?) - Display pool — anything live worth referencing - Domains from CLAUDE.local.md - Time since last activity ### 2. Show the snapshot Open with the snapshot, briefly. Example shapes: > *"Looks like you're mid-work on `` (formalization). 3 claims emitted, 2 verified, 1 still pending. `show_research` is live; verifier loop is running. Last activity ~12 minutes ago."* > *"No active session — clean slate. `CLAUDE.local.md` lists physics + math as your domains."* > *"You started `` 2 hours ago, set phase to exploration, then went quiet. No claims, no established results yet. Display pool has a `.html` from earlier."* Don't moralize — just describe what the state looks like. ### 3. Ask about their intent Open the conversation with the right question for the situation: - **No session** → *"What are you trying to do? Pick something to work on, or just talk through what's on your mind."* - **Active session, recent activity** → *"Still on ``, or want to shift?"* - **Stale session (> 4h since lastActive)** → *"You were on `` last time. Picking that up, or starting something new?"* - **Active session but stuck-looking** (work phase, no claims emitted in last 20 min) → *"How's the work going? Stuck somewhere, or just thinking?"* One question. Wait for the answer. ### 4. Identify position in the arc Based on the state + their answer, locate them in the five-step arc and surface it visually: ``` start → explore → [WORK] → check → wrap ▲ (you are here — 3 claims emitted, 1 unverified) ``` If the position is ambiguous (between two stages), say so and ask which fits better. ### 5. Recommend a concrete next move Based on position + state + their stated intent, suggest one specific next move with reasoning. Examples: - *"Two results just landed and you haven't `/check`-ed yet. Now's a good time — empty findings is the most valuable answer when the work is sound."* - *"You said you're stuck on the linear-algebra step. Worth `/explore`-ing alternative decompositions for 10 minutes, then back to `/work`?"* - *"You've been here 2 hours and you're winding down — `/wrap` saves the state for tomorrow. Takes 30 seconds."* - *"No goal yet. `/start` takes a minute and seeds the research state so the verifier has something to check against."* Always frame it as a suggestion the user can accept, modify, or override. Don't auto-invoke. ### 6. Offer a setup tweak (only if useful) Run a quick readiness check. Surface only the items that are *actually* off — don't lecture if everything's fine. | Check | Surface if | |---|---| | Domain in `CLAUDE.local.md` | The current goal touches a domain not in the user's profile (offer to add) | | MCP servers | Working in physics / math / ml but `sympy-mcp` / `desmos-mcp` not listed — point to `/skill-finder` or `claude mcp add` | | Verifier loop | They're emitting claims but `/loop /verifier` isn't running | | Watcher | They're actively working in the workspace panel but `/watcher` isn't running and they haven't declined it | | Display pool | More than 8 stale tabs (offer `cleanup_displays`) | | Research state | They're in formalization but no `init_research` ran (offer `/work` to seed it) | Surface at most two — don't overwhelm. Each as one line: *"Heads up: you're emitting claims but the verifier loop isn't running. Want me to start it?"* ### 7. End with the cheatsheet After orientation, show the five. This is the "if you want to do something specific, here's the menu": ``` /start Begin a session — set the goal, pick mode (explore vs work) /explore Generate ideas, survey, hypothesize — for "what should I ask?" /work Do the work — derive, prove, compute, analyze (verifier on) /check Push back at a milestone — challenge assumptions and conclusions /wrap Close out — summarize, persist for next session, clean up ``` *Plain-language phrasing also works — "let's derive X" → /work, "I'm done for today" → /wrap. You can always `/triptych` again to re-orient.* ## `/triptych setup` — readiness check only Skip the goal conversation. Run the full readiness check from step 6 and report: - ✓ items that are good (one line each, dim) - ⚠ items that could be better (one line each, with the fix as a suggestion) If everything's fine, just say so cleanly: *"Setup looks good for ``. No changes recommended."* ## `/triptych help` — static cheatsheet only The old behavior — no interaction. Just the five-command block plus a one-line description per command, plus the power-user direct skills list. Useful when the user explicitly wants the menu without a conversation. ## Power-user direct skills (Surface this section in `help` mode, or on request — not by default.) The five wrap most common needs. If you know exactly what you want, the underlying skills are still callable: - `/verifier`, `/watcher`, `/dashboard`, `/redteam`, `/whats-missing` — the spawned agents directly - `/scientific-method`, `/think-rigorously`, `/hypothesis-generation`, `/literature-review` — methodology - `/autonomous`, `/autoresearch` — special operating modes - `/physics-in-triptych`, `/math-in-triptych`, `/ml-in-triptych` — domain mentors - `/triptych-displays`, `/triptych-workspaces` — addon catalogs - `/skill-finder` — fetch skills from PRPM / K-Dense for sub-domains we don't bundle - `/field-report` — user-controlled feedback to the maintainer - Plus eight bundled K-Dense skills (`/scientific-writing`, `/sympy`, etc.) ## Hard rules - **Don't automate the decision.** `/triptych` *recommends*, never *executes* `/start` / `/work` / `/check` / `/wrap` automatically. The user picks. - **Read state before speaking.** Always show the snapshot first so the user knows what `/triptych` is reasoning from. - **One question at a time.** Don't interrogate. - **Surface setup gaps only when they matter.** No checklist read-aloud if everything's fine. - **Empty state is fine.** A new user typing `/triptych` immediately should get a clean orientation, not an error about missing session.