Datasets:
Nomad Migration Protocol
Defines how to bring an existing, scattered AI setup — a single custom-instructions box, a folder of prompt files, an existing single-assistant memory system, or an existing multi-agent setup that isn't yet Nomad-shaped — into the Nomad skeleton (see
team-skeleton.md) without losing information. This is Path B fromSKILL.md.
0. What this file assumes
Step 0 of SKILL.md already settled the language. This is the English version — see migration-protocol.zh-CN.md for the content-equivalent Chinese version. You already have (or are about to scaffold) a Nomad-shaped team root per team-skeleton.md; this file is about getting content into that skeleton from somewhere else, not about the skeleton's shape itself. Whether the target root is actually ready to receive content is checked in Phase 0a below, not assumed here.
1. Core principles (non-negotiable)
These come before any procedure below. If a step in this file and one of these principles ever seem to conflict, the principle wins.
- Preserve, don't summarize-rewrite. When the source already has a complete, coherent document (a personality file, a preference list), copy it forward largely verbatim. Only restructure when the old content genuinely can't map onto a Nomad slot — and when you do restructure, keep every original detail; compressing "for tidiness" is a fidelity violation, not a cleanup.
- Inventory before asking, ask before writing. Scan first. Show the user what you found. Get their go-ahead. Zero files get created or modified before that go-ahead is given.
- Report every category, including the empty ones. "Found nothing" is a valid, reportable result — it is never a reason to omit a category from the final report. Never fabricate content to fill a category that's genuinely empty.
- When classification is unclear, ask — don't guess. Whether a piece of content is "personal preference" vs. "project knowledge," or which agent should own a given skill, is not always obvious from the content alone. A wrong silent guess is worse than a short clarifying question. Proviso — when there's no one to ask (an unattended/automated run, or the user's pre-supplied answers don't cover this case): take the most conservative option available — the one that loses the least information and commits to the least — mark the item
OPEN QUESTIONdirectly in the landed content, and list it in its own section of the Phase 7 reconciliation statement rather than burying it in a regular row. This is not a license to skip asking whenever asking is merely inconvenient; it only applies when asking is genuinely not possible in the moment. - Never silently drop anything. Anything not confidently placed in this pass goes to a labeled staging area (§6) for the user to resolve later — it does not vanish, and it is not left half-imported without a note explaining why.
2. The five information types
Everything found during inventory sorts into one of five types. This is the classification vocabulary used for the rest of this file.
| Type | What it covers |
|---|---|
| Personality | Core identity, mission, values, voice/tone, behavioral rules — whatever makes "this AI" recognizably itself |
| Memory | Structured long-term knowledge: preferences, decisions, working experience, notable events |
| Skills | Reusable methods, workflows, templates, playbooks — knowledge about how to do something, not about the user |
| Project knowledge | Context tied to a specific ongoing or completed piece of work |
| Session history | Raw, unedited records of past conversations — the "ore," not the "refined metal" |
Memory vs. Project knowledge — the recurring judgment call. "Decision" content is the one item that plausibly fits either type, and re-litigating it on every migration wastes time. The dividing line: a decision bound to one specific, identifiable project belongs in Project knowledge (lands in that project's projects/{name}/ decision log, per §4); a decision that's personal or applies across projects — a working-style choice, a standing preference, a general policy — belongs in Memory. When a decision genuinely spans both (a project-specific choice that also set a general precedent), file it under Project knowledge and add a one-line cross-reference from Memory rather than duplicating it in full.
3. The three hardest types to lose
These rarely live in any single file, so a naive file-by-file scan misses them. Ask about them explicitly during inventory — don't wait for the user to volunteer them.
- Tone and values that were never written down as a rule. "Was there a particular register your prior AI used with you — direct vs. gentle, formal vs. casual, any running jokes or ways it addressed you — that you'd want carried forward?"
- Unwritten tacit agreements. "Are there things you and your prior setup just came to 'know' without ever writing it down — a habit it picked up, something you stopped having to explain?"
- Implicit relationships between files. "Do any of your files only make sense next to another one — a memory note that references a project file informally, a shorthand that assumes context from somewhere else? Point them out so the mapping keeps the link intact."
4. Procedure
Phase 0 — Identify the source
Ask, or infer: is the source a single persona (one file, one custom-instructions box, one export) or an existing multi-agent setup? This determines whether everything found lands under one {AGENT_NAME}/ home or gets split across several.
Before moving to Phase 1, settle a concrete name for every {AGENT_NAME}/ home this migration will land content into — reuse the source's existing name if it had one, or pick one now (see team-skeleton.md §7 design lesson 3 for naming-theme guidance). {AGENT_NAME} is not something that emerges naturally partway through Phase 4's landing map — every row in that map depends on it, so it has to be fixed before any writing starts, not discovered along the way.
Phase 0a — Target readiness check
Before starting inventory, confirm the target team root actually meets team-skeleton.md §8's minimum-viable-team bar. If it doesn't yet, walk the user through the minimum Path A scaffold first (team-config.json + path-conventions.md + at least one agent home's four minimum pieces + an empty team-log.md), then come back here — migration is about landing content into a skeleton that already exists, not about creating the skeleton as a side effect.
Phase 1 — Inventory
Scan every location the user points to (and ask "is there anything I can't see — local files, another platform's export, anything on paper?" — you cannot inventory what you were never told about). For each of the five types, produce a count and a location list, even when the count is zero. Copy the raw originals, unmodified, into a staging buffer (see §6) before doing anything else — this is your rollback point if something goes wrong downstream.
Phase 2 — Dry-run preview
Before writing anything to a real Nomad slot, produce a full mapping table and show it to the user:
| Source item | Type | Target slot | Action | Notes |
|---|---|---|---|---|
| (one row per item found) | create / append / merge |
This is a hard gate, not a formality. After presenting the complete preview table, STOP and wait for the user's explicit approval. Before that approval arrives, creating or modifying any target file is prohibited — this includes writing a "draft" version ahead of time "for efficiency," staging a file that only needs a final rename, or any other pre-write that happens before the user has actually said yes. Zero real writes happen at this stage, full stop. This table is the user's chance to catch a wrong classification before it's permanent.
Known failure mode. In real-run testing, the executing AI has been observed writing target files first and only presenting the mapping table afterward, as a post-hoc summary rather than a genuine approval gate — this is the single most commonly broken rule in this entire protocol. If you notice yourself reaching for a write tool before the user has responded to the preview table, stop; that impulse is the failure mode, not an exception to it.
Phase 3 — User selects scope
The user may approve the whole preview, or select a subset by type ("just personality and memory for now, skip skills"). Partial import is a first-class option, not a fallback — don't push for "just do it all" if the user wants to go incrementally.
Phase 4 — Import per type (the landing map)
Execute only the approved rows, using this mapping into the team-skeleton.md structure:
| Type | Nomad landing slot | Notes |
|---|---|---|
| Personality | {AGENT_NAME}/AGENTS.md (generate the CLAUDE.md thin shell alongside it if this deployment will use Claude Code — see team-skeleton.md §2) |
Verbatim copy wherever the source already has a coherent personality doc. Content that doesn't fit any template field goes in an "Other" section at the end — never dropped for not fitting the mold. |
| Memory | {AGENT_NAME}/memory/preferences.md (create it — accumulating imported memory content is itself the trigger condition for building this file, per team-skeleton.md §2's optional-pieces table) |
Preferences and decisions verbatim; experience may be condensed but never stripped of concrete parameters (commands, paths, numbers); events may be summarized |
| Skills | Reusable across every agent → team root 90-methodology/. Deep, agent-specific methodology → that agent's memory/playbook.md. |
When it's not obvious which bucket a skill belongs in, ask — don't default to one bucket silently. When the team currently has only one agent, "reusable across every agent" and "specific to this agent" are both trivially true — headcount can't decide the split. Judge it by portability instead: if a future new agent joining the team should automatically inherit this skill, it belongs in team root 90-methodology/; if not, it stays in this agent's own memory/playbook.md. |
| Project knowledge | Owned by one identifiable agent, still active → that agent's projects/{project-name}/ (see team-skeleton.md §2 — this is a persistent slot, distinct from workspace/, which stays scratch-only); owned by one agent, mid-handoff → that agent's memory/handoff.md; relevant across multiple agents → team root 02-shared-knowledge/ |
The skeleton deliberately has no single team-root-level "projects" folder the way a single-persona memory system might — project ownership in a team is distributed across whichever agent's domain it falls under, each in its own projects/ slot. Ask the user which agent should own a given project when it isn't obvious. |
| Session history | Full, unedited transcripts → {TEAM_ROOT}/98-archive/sessions/{AGENT_NAME}/YYYY-MM/. Optionally add one summary pointer line in that agent's memory/worklog.md. |
The archived transcript is the record of truth; the worklog pointer is a convenience index, not a substitute. Never import only a summary and report it as "session history imported." |
Phase 5 — Conflict resolution
A conflict is any case where the target slot already has real (non-placeholder) content. For each conflict, offer a three-way choice:
- Skip — leave the existing content untouched, drop the incoming item
- Overwrite — replace the existing content with the incoming item (only for genuinely superseding content — confirm this is what the user means, since Nomad's memory files are normally append-only per
team-skeleton.md§6) - Rename — keep both; the incoming item lands under a disambiguated name/section
If several conflicts are clearly the same kind (e.g., re-running an import that partially succeeded before), the user may set one rule for "all remaining conflicts in this batch" instead of confirming one at a time — but the default is per-conflict confirmation, not batch silence.
Phase 5b — Source-internal contradictions
Phase 5 covers conflicts between an incoming item and existing target content. A different kind of conflict can also surface during Phase 1 inventory: two or more source items contradict each other — a disagreement that predates this migration and has nothing to do with the target slot. Do not silently resolve this by picking one and discarding the other; that destroys information the source itself never resolved.
For each source-internal contradiction found, offer the user a three-way choice:
- Keep both, annotated — land both versions in the target slot with an inline note flagging the disagreement and its sources
- Prefer the newer, annotated — if the items carry dates, land the newer one as primary and the older one as an annotated note beneath it (not deleted)
- Ask the user to decide — when neither of the above is clearly right, put the question to the user directly
Default when the user can't be reached in the moment (see §1 principle 4's proviso): if the contradicting items carry dates, the newer one wins by default — but both must be preserved in the record, never overwrite-and-discard. This is the same non-negotiable as principle 1 (preserve, don't rewrite) applied to the source itself, not just the target.
Phase 6 — Staging and undecided items
Reuse the existing 97-inbox-misc/ folder (see team-skeleton.md §1) rather than inventing a new top-level location:
97-inbox-misc/migration-{YYYYMMDD}/raw/— unmodified copies of every source item, made during Phase 1, before any classification write happens. This is the rollback point.97-inbox-misc/migration-{YYYYMMDD}/undecided/— anything that, even after asking the user, still couldn't be confidently placed. Every item here gets a one-line companion note explaining why it's stuck (ambiguous ownership, user deferred the decision, format didn't fit any of the five types). Nothing sits here silently — an entry with no reason attached is itself a bug in the migration run.
Phase 7 — Reconciliation statement
Close every migration with a statement covering all five types, including the ones with zero items:
| # | Original source | Type | New location | Fidelity | Notes |
|---|---|---|---|---|---|
| 1 | (file/section) | Personality | 01-agents/{name}/AGENTS.md |
Verbatim | — |
| 2 | (file/section) | Memory | 01-agents/{name}/memory/preferences.md |
Verbatim | — |
| — | (none found) | Session history | — | — | 0 items — user confirmed no prior session logs exist |
Report format for the categories with zero items: state the count as zero explicitly ("Skills: 0 items — none found, user confirmed nothing to migrate") — never omit a category from the statement because it happened to be empty.
Open questions get their own section. Any item landed under §1 principle 4's "no one to ask" proviso is OPEN QUESTION-tagged at the landing site and listed again here, in a section separate from the five-type table above — don't fold it into a regular row where it can get lost.
5. Fidelity discipline
Priority order when the four goals below pull in different directions (highest first):
- Complete information retention over tidy formatting. When in doubt, keep more, not less.
- Copy first, organize second. Don't cherry-pick during the copy pass — that's how information quietly disappears. Get everything into the staging buffer, then classify.
- Format is a tool, not a cage. The landing map in §4 is the default shape. If a piece of content genuinely doesn't fit any slot cleanly, say so and ask, rather than forcing a bad fit.
- A restructuring pass must flag its own compression ratio. If turning source material into a landing-map file drops more than roughly half the original length, say so explicitly and let the user decide whether that's acceptable — don't silently ship a lossy conversion as if it were complete.
6. Sensitive content during migration
Two different situations, handled differently:
- Already present in the source the user pointed you to (credentials, personal contact details, addresses, financial figures already sitting in their old memory files) — this was already something the user previously entrusted to their prior setup. It may be migrated forward with a note ("carried over from prior setup, not newly collected"); it does not need to be re-confirmed item by item.
- Newly surfacing during migration (something you notice only because you're now reading files you weren't reading before, or the user mentions something new in the course of answering an inventory question) — treat this the same as any other newly-discovered sensitive information in normal operation: confirm with the user before writing it anywhere, rather than assuming prior consent extends to it.
7. Minimum bar for "migration complete"
A migration run is not done until:
- Phase 1's inventory counts (including zeros) have been shown to the user
- Phase 2's dry-run table was shown and approved before any real write
- Every approved item has landed in its target slot per §4, or is logged in
97-inbox-misc/migration-{date}/undecided/with a reason - Phase 7's reconciliation statement has been delivered, covering all five types
Skipping straight from "scan" to "done, everything's imported" without steps 2–4 is not a faster migration — it's an unverified one.
Nomad skill · migration-protocol reference · desensitized — no real paths, names, or account details appear anywhere in this file. Version 0.3 · Last updated 2026-07-14