Guided first journey into Trinity — one command that takes a newcomer from "what is Trinity?" through choosing their path, installing or connecting an instance, wiring the MCP connection, and getting a first agent alive — with a live smoke test at every step. A resumable concierge, not a manual — each stage hands off to the specialist skill that owns it (connect, deploy-new-instance, onboard, the create-agent wizards), and platform questions are always answered from live documentation, never static text — before any instance exists via the public Trinity Docs Q&A endpoint (Vertex AI Search over docs/user-docs, resynced on every release) and the user-docs index on GitHub, after connecting via the instance's own ask_trinity MCP tool.
Instrucciones de origen · Vista previa de solo lectura
name
start-here
description
Guided first journey into Trinity — one command that takes a newcomer from "what is Trinity?" through choosing their path, installing or connecting an instance, wiring the MCP connection, and getting a first agent alive — with a live smoke test at every step. A resumable concierge, not a manual — each stage hands off to the specialist skill that owns it (connect, deploy-new-instance, onboard, the create-agent wizards), and platform questions are always answered from live documentation, never static text — before any instance exists via the public Trinity Docs Q&A endpoint (Vertex AI Search over docs/user-docs, resynced on every release) and the user-docs index on GitHub, after connecting via the instance's own ask_trinity MCP tool.
{"version":"1.3","created":"2026-08-06T00:00:00.000Z","author":"Ability.ai","changelog":["1.3: Smoke test no longer treats an empty agent list as the fresh-install signature — ent#124 seeds the acme trio plus Cornelius, so a fresh instance boots ~4 agents (Stage 4 already assumed seeded agents existed)","1.2: Stage 4 teaches the repository-first deploy sequence — push the agent to GitHub and add the instance's GitHub token (Settings → GitHub token) before /trinity:onboard, which then deploys by cloning the repo; the local-file deploy is named as the fallback that offers promotion afterwards","1.1: Ground the whole journey in live docs — pre-connect questions now go to the public Trinity Docs Q&A endpoint (Vertex AI Search over docs/user-docs, no auth, no instance needed) and Stage 0 orients against the live user-docs index on GitHub; the static narrative is demoted to a fallback for when the network is down","1.0: Initial version — resumable five-stage guided journey (orient → choose your door → get an instance → connect MCP + smoke test → first agent alive), routing every operational step to the specialist skill that owns it and using ask_trinity as the live documentation channel once connected"]}
Trinity — Start Here
ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. start-here vX.Y — recent: <summary>. Then proceed.
The front door to Trinity. One command walks a newcomer from "what is this?" to "my first agent answered me from my own instance" — and costs nothing until they explicitly choose a step that deploys something.
Design rules (bind every stage):
Route, don't duplicate. Every operational step belongs to a specialist skill — /trinity:connect, /trinity:deploy-new-instance, /trinity:onboard, /create-agent:create, /agent-dev:agent-fleet-analysis. This skill sequences them and carries the user's state; it never re-implements their flows.
Live documentation over static claims — at every stage. Never improvise platform facts from memory. Two grounded channels cover the whole journey (see Live documentation channels below): before any instance exists, the public Trinity Docs Q&A endpoint and the live docs/user-docs index on GitHub; after connecting, the instance's own mcp__trinity__ask_trinity. The short Stage 0 block is a network-down fallback, not the source of truth.
Zero commitment until chosen. The tour door deploys nothing and asks for no credentials. Stages that create or deploy things happen only after the user picks them.
Always resumable. State persists across sessions; re-running this skill continues where the user left off.
Live documentation channels
Pre-connect (no instance, no auth, works from any laptop):
Trinity Docs Q&A — a public endpoint backed by Vertex AI Search over the platform's docs/user-docs/**, onboarding docs, and the Trinity Compatible Agent Guide, re-indexed automatically on every push to the platform's main branch — always current:
Response: {"answer": "...", "state": "SUCCEEDED", "session_id": "..."}. Contract notes: pass session_id back for multi-turn follow-ups, but treat it as an opaque string (it exceeds 2^53 — numeric handling corrupts it); sessions expire silently (~30 min) — if the returned session_id differs from the one you sent, context was lost, mention nothing and carry on; errors come back as {"error": "..."} or, from Google's frontend, as HTML.
The user-docs index — fetch https://raw.githubusercontent.com/abilityai/trinity/main/docs/user-docs/README.md to see the current documentation map, and deep-link the user to specific pages as https://github.com/abilityai/trinity/blob/main/docs/user-docs/<path>.
Post-connect: prefer mcp__trinity__ask_trinity (the instance's own docs tool — same grounded corpus, plus it lives where the user's fleet lives). The public endpoint remains the fallback whenever the instance is unreachable.
Channel routine: any time the user asks anything about Trinity — at any stage — answer through whichever channel is available, quote the grounded answer, and offer the relevant user-docs deep link when one exists.
Journey State
State lives at ~/.trinity/start-here.json (same home as the connection profile):
If the argument is reset: delete the file, start from Stage 0.
If state exists and stage > 0: print a one-line recap ("You're at stage N — last time you {summary}") and ask: continue where you left off, or start over?
Otherwise: begin at Stage 0.
After every stage: write the updated state back (stage reached, door chosen, completed entries appended, ISO timestamp). Never skip the write — the resume promise depends on it.
STAGE 0: Orient — what Trinity is
Print this orientation, then stop narrating (rule 2):
## What is Trinity?
Trinity runs Claude Code agents as an always-on fleet.
- An agent is a folder: CLAUDE.md (identity) + skills + memory files,
driven by the same Claude Code harness you're using right now.
- The bet: the agent behaves identically on your laptop and on Trinity.
You build and test locally; Trinity makes it durable — schedules,
Slack/Telegram/email channels, dashboards, a fleet you can talk to.
- This session becomes the control plane: once connected over MCP, you
list, message, and manage your whole fleet from right here.
This tour is free — nothing gets deployed until you choose it.
Then orient against the live docs (rule 2): fetch the user-docs index (channel 2 above) and add one live line under the block — the newest release from its What's New section with the deep link, e.g. "Current release: v<A.B> — what's new: " (always the version the fetch returned — never a remembered one). If the fetch fails, skip this line silently; the static block stands alone.
Then ask (AskUserQuestion) whether they have questions before moving on. Answer every question through the public Docs Q&A endpoint (multi-turn — keep the session_id), quoting the grounded answer with a deep link where one fits. There is no "wait until you're connected" — the docs are live from the first minute.
→ Write state (stage: 1), go to Stage 1.
STAGE 1: Choose your door
Use AskUserQuestion — Question: "Where are you right now?" — Header: "Your path"
Just looking — show me around, deploy nothing
Build my first agent — I know roughly what I want it to do
I already have agents — Claude Code folders, n8n flows, or framework apps
Set up my instance — I'm ready to install or connect Trinity
Door behavior:
Just looking → give the extended tour: walk one concrete day-in-the-life (an agent that wakes on a schedule, reads its inbox channel, does its job, reports to a dashboard). If an MCP connection already exists, make it live: mcp__trinity__list_agents + mcp__trinity__get_fleet_health and narrate what's actually running. Then re-offer the other three doors. Record door: "tour".
Build my first agent → the create-agent wizards own this. If the plugin isn't installed, have them run /plugin install create-agent@abilityai first, then /create-agent:create. Tell them plainly: the agent will work locally first — Trinity is the upgrade, not the gate. Record door: "build" and agent_dir once created; when the wizard finishes, resume here → Stage 2.
I already have agents → /plugin install agent-dev@abilityai, then /agent-dev:agent-fleet-analysis pointed at their agents' parent directory — a read-only maturity report they get value from before deploying anything. Record door: "bring"; when they're ready to host the fleet → Stage 2.
Set up my instance → record door: "install", go straight to Stage 2.
STAGE 2: Get an instance
Ask: "Do you already have a Trinity instance URL?"
Yes → note instance_url, skip to Stage 3.
No → /trinity:deploy-new-instance owns provisioning; it will ask them to choose between cloud (ability.ai — managed, no infrastructure), a self-hosted server, or local Docker on this machine. Run it, and when an instance is reachable, return here → Stage 3.
Not yet / just exploring → perfectly fine. Save state and close warmly: "Your progress is saved — run /trinity:start-here whenever you're ready and we continue at this exact step."
STAGE 3: Connect the MCP — and prove it works
This is the moment the session becomes a control plane.
Connect: run /trinity:connect — it is the single writer of .mcp.json (email OTP → MCP API key → config). Do not hand-roll any part of its flow, including its error handling (e.g. the no-code-arrives whitelist case — the connect skill explains it).
Load the tools: if mcp__trinity__* tools are not available in this session after connect, that's expected — the MCP server loads on restart. Tell the user: restart Claude Code (or run /mcp to reconnect), then run /trinity:start-here again — state resumes exactly here. Write state (stage: 3) before saying this.
Smoke test — run all three, report a checklist:
Check
Tool
Pass looks like
Fleet visible
mcp__trinity__list_agents
A list returns. A fresh install is not empty — it seeds the acme system (scout, sage, scribe) plus Cornelius, so expect ~4 agents (ent#124). An empty list is still a pass; it just means the seed was disabled or this instance isn't fresh
Instance healthy
mcp__trinity__get_fleet_health
Health summary returns, no critical agents
Live docs online
mcp__trinity__ask_trinity
Grounded answer returns
For the ask_trinity check, ask something the user actually wants to know (or default to "What can agents do on a schedule?"), show the answer, and tell them explicitly: this is the documentation channel from now on — any question about how Trinity works gets a live, doc-backed answer from your own instance.
Print the checklist with PASS/FAIL per row. On any FAIL, fix before proceeding: unreachable instance → verify it's up (their ops agent's /status if they deployed one); auth issues → /trinity:connect --force.
→ Write state (stage: 4, completed += "mcp-connected").
STAGE 4: A first agent, alive
Goal — the user exchanges a real message with an agent running on their instance. Branch on their situation:
They built an agent locally (door "build"): deployment is repository-first, so set that up before deploying — it's two small steps and it's how every later update reaches the agent:
Push the agent to GitHub (gh repo create <name> --private --source=. --push from its directory) — Trinity deploys by cloning the repo, so the repo is the deliverable.
Add a GitHub token in the instance UI under Settings → GitHub token (a fine-grained PAT, Contents: Read) — required for private repos, recommended for public ones.
Then, from the agent's directory, /trinity:onboard deploys it from that repo and reports which commit landed. (No repo, or GitHub unreachable? onboard falls back to a local-file deploy — fine to get moving, and it offers to put the agent on the repo path afterwards.) Then, from here, mcp__trinity__chat_with_agent — send a short real task, show the reply.
The instance already has agents (seeded fleet or door "bring" after migration): pick one from list_agents and chat_with_agent it with a hello-task.
Neither yet: offer /create-agent:create now, or chat_with_agent against any seeded agent just to feel the loop.
Practical note: if a chat_with_agent call returns a queued_timeout receipt, the task is running — poll mcp__trinity__get_execution_result with the returned execution_id; never blind-retry (it would duplicate-queue).
→ Write state (stage: 5, completed += "first-agent-conversation").
STAGE 5: Wrap — and where each road leads
Print a recap of what they now have (checklist built from completed), then the hand-offs — each is a door out, not more tour:
## You're running. Where to next?
- Another agent → /create-agent:create (wizard menu)
- Bring your existing bots → /agent-dev:agent-fleet-analysis
- Design an agentic system → describe the job to ask_trinity, then
/create-agent:custom or :kb-agent for the shape
- Day-2 operations → /trinity:sync (local↔remote),
/trinity:loop (bounded remote loops),
schedules & channels: just ask — the live
docs answer via ask_trinity
- Docs assistant anywhere → claude mcp add trinity-docs -- npx -y @abilityai/trinity-docs-mcp
(the same grounded Q&A in any session, no
instance required)
Mark state stage: 5, completed += "journey-complete". Re-running the skill later greets them as a returning user and offers the hand-off menu directly.
Error Handling
Situation
Response
State file corrupt/unreadable
Say so, offer reset — never crash the journey over bookkeeping
Specialist skill not installed
Give the exact /plugin install <name>@abilityai line, then continue
MCP tools absent after connect
Normal — restart Claude Code or /mcp, re-run /trinity:start-here, resumes at Stage 3
Instance unreachable mid-journey
Point at their ops agent (/status) if one exists, else /trinity:connect --force; save state first
User wants to stop anywhere
Save state, confirm resume works, close warmly — no guilt-tripping
Notes for maintainers
Never grow the Stage 0 narrative. It is deliberately a network-down fallback; every other platform fact must come from the live channels at run time. The drift surface to review when the platform changes is exactly three things: the Stage 0 block, the Docs Q&A endpoint URL, and the user-docs raw/deep-link URLs — the docs content behind them keeps itself current (resynced to Vertex on every push to main).
This skill's hand-off targets are contracts: connect (single .mcp.json writer), deploy-new-instance, onboard, create-agent:create, agent-dev:agent-fleet-analysis. If any of those are renamed or resharded, update the routes here in the same change.
