---
name: nomad
description: Turn any folder into a portable, multi-harness AI team workspace — build a new team from scratch, migrate an existing scattered setup into a clean skeleton, or mount an already-built team onto a new AI harness (Claude Code, Codex/ChatGPT, OpenCode, Copilot, Cursor). Trigger on requests like "make my AI team portable", "give my agents persistent memory across tools", "set up a multi-agent team folder", "migrate my prompts/config into a team structure", "connect this team to another AI tool", "sync my agent across tools", "AI team memory portability", or "switch AI tools without losing my agent's memory".
---

# Nomad — Portable AI Team Workspace

Nomad teaches an AI agent to turn a plain folder into a **portable, multi-harness team workspace**: a folder that, when opened by any capable AI tool, self-configures that tool with the right personality, memory, and collaboration rules — no cloud service, no proprietary format, just files.

This is not a product to install. It is a **skill**: read it, then act on the user's folder directly.

**Terminology note**: the generic term for "an AI tool/CLI/IDE capable of reading files and running an agent session" is **harness** throughout this skill and its reference files — Claude Code, Codex/ChatGPT, OpenCode, Copilot, and Cursor are all harnesses. **Platform** is reserved for passages describing one *specific* harness's own behavior — e.g. the per-platform table in `mounting-guide.md` §3. If the two appear used loosely elsewhere in these files, treat them as referring to the same concept.

## Step 0 — Ask the user's language, always

Before doing anything else, ask (or infer from how the user is writing to you):

> "Would you like to proceed in English or 中文?"

Every reference file in this skill has two parallel versions — `references/<name>.md` (English) and `references/<name>.zh-CN.md` (Chinese). Once the language is known, **read only the matching version** for the rest of this task. The two versions are content-equivalent; never mix languages mid-task, and never presume English by default just because this file is in English.

## Step 1 — Figure out which of the three paths the user needs

Ask, or infer from context, which of these three the user actually wants. They are independent — a user may want only one.

| Path | When to use it | What to read next |
|---|---|---|
| **A. Build a new team from scratch** | The user has an empty folder (or nothing yet) and wants to stand up a new multi-agent team with portable structure from day one. | `references/team-skeleton.md` (or `.zh-CN.md`) |
| **B. Migrate an existing setup** | The user already has agent configs, prompts, memory files, or a messy team folder and wants it imported into the Nomad skeleton without losing anything. | `references/migration-protocol.md` (or `.zh-CN.md`) |
| **C. Mount a built team onto a new harness** | The team already has Nomad-shaped structure and the user wants to connect it to one more AI tool (add Claude Code alongside Codex, etc.). | `references/mounting-guide.md` (or `.zh-CN.md`) |

Do not guess silently when the request is ambiguous — a one-line clarifying question is cheap; building the wrong thing is not.

Before committing to Path A on the user's say-so alone, glance at the target directory. If it already has agent configs, prompt files, or memory-shaped content — even though the user described it as "starting fresh" — stop and confirm: "this folder already has content in it; do you want to keep it and migrate (Path B), or confirm you want to wipe/ignore it and scaffold clean (Path A)?" Don't silently trust a verbal "empty folder" description over what's actually on disk.

## Step 2 — Execute per path

- **Path A**: Open `references/team-skeleton.md` (pick the language version per Step 0). It defines the full folder structure, the parameterization mechanism (`team-config.json`), the markdown-layer/node-layer split, the collaboration protocol, and a set of hard-won design lessons — plus, as of v0.4, three governance layers for teams that outgrow the basics: closed-set memory types with a slot-application channel (§9), a file-lifecycle policy (§10), and an optional metadata/relations layer (§11). Follow it to scaffold the user's folder. Resolve **every** field in the `team-config.json` schema (`team-skeleton.md` §3) rather than a partial list backfilled with silent defaults — for most fields (team root path, principal's name/language, `os`, naming style, notification channel) that means asking the user directly; for `confirmation_gates` specifically, follow the gate procedure in the Ground rules below (conservative default first, not a blank "what do you want" question) instead of asking it the same way as the other fields. Note: the **node-layer scripts now ship in this package's `scripts/` directory** (see `team-skeleton.md` §4 for how the md/node layers relate, and `scripts/README.md` for setup and wiring instructions) — the md layer remains fully specified and buildable on its own; the node layer is optional acceleration on top.
- **Path B**: Open `references/migration-protocol.md` (pick the language version per Step 0). It defines the readiness check → inventory → dry-run preview → per-type import → conflict resolution (including source-internal contradictions) → reconciliation-statement procedure for bringing an existing scattered setup into the skeleton without losing information. Requires `team-skeleton.md` to already be scaffolded or in progress (Path A), since migration is about landing content into that structure.
- **Path C**: Open `references/mounting-guide.md` (pick the language version per Step 0). It defines the per-platform connection steps, known pitfalls, and a post-mount acceptance checklist for adding one more AI harness to an already Nomad-shaped team. Requires the team to already meet the minimum-viable-team bar in `team-skeleton.md` §8.

## Ground rules that apply to every path

- **Never hardcode the user's real paths, name, or risk preferences into anything, full stop — not just things you intend to hand off as a reusable template.** This applies even to a one-off deployment that will only ever serve this one user: hardcoding forecloses the migration/mounting/multi-machine paths this skill exists to support. Everything that varies by user goes into `team-config.json` or an explicit placeholder — see `references/team-skeleton.md` §3 and the fourteen hardcoding points in §5.
- **The confirmation-gate tier follows one procedure, every time: give the conservative default first → tell the user what that default is and ask if they want to loosen any of it → only loosen what they explicitly confirm.** Never skip straight to "just ask the user what they want" (that risks silently inventing a posture along the way) and never ship a pre-loosened gate as if it were the template standard. See `references/team-skeleton.md` §7 design lesson 2 for the full gate-list mechanism this procedure produces.
- **Templates keep placeholders; deployments resolve them.** Anything you are handing off as a reusable template (or that ships as part of this skill) keeps its placeholder markers (`{PLACEHOLDER_NAME}` etc.) literally in place. Anything you are actually standing up for a real user gets those placeholders resolved to real values before you call it done — a file that still shows a placeholder where a real value belongs is not a finished deployment.
- **Before writing to any target location, check what's already there.** A location that already holds real (non-placeholder) content is a conflict, not a green light to overwrite — resolve it with the skip/overwrite/rename choice in `references/migration-protocol.md`'s Phase 5, even if you arrived here via Path A rather than Path B.
- **A dry-run preview is a hard stop, not a formality.** For Path B specifically (see `references/migration-protocol.md` Phase 2), after showing the mapping table, wait for the user's explicit approval before creating or modifying any target file — no pre-written drafts "for efficiency." This is the single most commonly broken rule observed in real runs of this skill; treat any urge to write before approval arrives as the failure mode itself, not an exception to it.
- **Don't claim a step is done if you didn't verify it.** If you scaffolded hooks config but didn't actually test that the hook fires, say so — untested wiring is not the same as working wiring.
- **This skill is documentation-first.** The `references/` files are the source of truth; this file is only a router. If this file and a reference file conflict, the reference file wins — update this file instead of trusting your memory of it.

Once Path A's md layer is scaffolded and working on its own, wiring the optional node-layer acceleration (hooks for session-start/stop/precompact, plus the concurrency-safe append tool) is a natural next step — see `scripts/README.md`.

**Governance artifacts (v0.5)**: when a team grows into the v0.4 governance layers (`team-skeleton.md` §9–§11), don't write the governance files from scratch — `references/governance/` ships a ready-to-use team-structure charter template and a metadata-conventions manual (bilingual pairs), and `scripts/` ships a four-check structure sentinel plus an assembled-team log-append variant (see `scripts/README.md` §7). Adapt the templates to the team's real slots instead of inventing a parallel scheme.

---
*Version 0.5 · Last updated 2026-07-23*