pmm-onboard

name	pmm:onboard
description	Seed your user identity layer from a prior AI or direct interview. Generates user.md (operative identity — how you think, communicate, decide) and routes PII to secrets.md. Works solo or in org-wide deployments (gated to a single designated agent).
argument-hint	[extract|interview|refresh]

pmm:onboard

Seed the user identity layer in PMM. Produces ${PMM_MEMORY_ROOT:-memory}/user.md — a structured identity file that tells the AI who it's working with and how to operate.

When to run:

• First time using PMM (after pmm:init)
• Switching from another AI (ChatGPT, Gemini, Copilot, etc.)
• Refreshing an existing identity layer after significant change

What it produces:

• ${PMM_MEMORY_ROOT:-memory}/user.md — operative identity (no PII). Committed to git.
• ${PMM_MEMORY_ROOT:-memory}/secrets.md — PII entries (name, location, people, employer). Gitignored.

---

Org-Wide Gating

In org-wide deployments, this skill is restricted to the designated onboard agent.

Before running, read ${PMM_MEMORY_ROOT:-memory}/config.md and check for:

onboard_agent: <handle>

If this field exists and the current agent is NOT the designated handle: refuse and redirect.

This skill is restricted to the <handle> agent in this deployment. Ask the coordinator to run vera:intake instead.

If the field does not exist (solo PMM user): proceed normally.

---

Argument Parsing

Parse $ARGUMENTS:

Argument	Mode
(no args) or extract	Extract mode — prompt the user to run the extraction interview in their old AI
interview	Interview mode — direct Q&A, no prior AI needed
refresh	Refresh mode — re-run against existing user.md, diff and confirm

---

Mode 1: Extract (default)

The user has a prior AI (ChatGPT, Gemini, Copilot, etc.) that already knows them. Extract that knowledge rather than re-interviewing from scratch.

Step 1 — Present the extraction prompt

Tell the user:

We're going to get your current AI to summarise everything it knows about you — rather than me asking you a bunch of questions. Copy the prompt below and paste it into [their AI]. Then paste the output back here and I'll build your identity files from it.

Read ${CLAUDE_PLUGIN_ROOT}/references/onboard-extraction-prompt.md and present the full extraction prompt for the user to copy.

Step 2 — Receive and parse the output

When the user pastes the output back:

1. Read it carefully. Note what's rich vs. what's thin.
2. Identify content for each user.md section (see Section Mapping below).
3. Identify PII that must route to secrets.md (see PII Routing below).
4. Check for gaps — if critical sections are missing, ask targeted follow-up questions (max 3 questions). Do not run a full interview if most content is covered.

Step 3 — Generate files

user.md — read ${CLAUDE_PLUGIN_ROOT}/references/templates.md for the user.md template. Fill each section with operative content only. Every sentence should change how the AI behaves — if it's descriptive but not operative, cut it.

secrets.md — append PII entries under appropriate section headers. Do not overwrite existing secrets.md content (it may contain credentials).

Step 4 — Confirm and commit

Show the user what will be written to each file. Wait for confirmation.

git add ${PMM_MEMORY_ROOT:-memory}/user.md && git commit -m "memory: onboard — user identity layer seeded"

secrets.md is gitignored — no commit needed for PII entries.

Step 5 — Report

User identity seeded.

• user.md: [N sections populated]
• secrets.md: [N PII entries routed]

Run pmm:onboard refresh any time to update.

---

Mode 2: Interview

No prior AI to extract from. Direct Q&A — targeted, not exhaustive.

Step 1 — Ask targeted questions

Use the AskUserQuestion tool for each. Keep it to 5-7 questions max.

Q1: Identity basics

What should I call you? Where are you based? Any language preferences (UK/US English, etc.)?

Q2: Communication style

How do you like AI to talk to you? Direct or diplomatic? Short or detailed? Anything that makes a response feel immediately wrong?

Q3: Working style

How do you process information and make decisions? Any personality type you identify with (MBTI, DISC, or just a description)?

Q4: Anti-patterns

What should AI never do when working with you? Things that annoy you, patterns you've had to correct before?

Q5: Operating modes

When you're brainstorming vs. making a decision vs. asking for feedback — do you want the AI to behave differently in each case? How?

Q6: Principles (optional — skip if user seems impatient)

Any beliefs or principles you'd want the AI to actually internalise — not just follow as rules, but understand as your operating system?

Q7: Key people (optional)

Are there specific people I'll encounter often? Names, roles, how you work with them?

Step 2 — Generate files

Same as Extract Mode Step 3. Route PII to secrets.md, operative content to user.md.

Step 3 — Confirm and commit

Same as Extract Mode Steps 4–5.

---

Mode 3: Refresh

User already has user.md and wants to update it.

Step 1 — Read current state

Read ${PMM_MEMORY_ROOT:-memory}/user.md. Present a summary of what's currently in each section.

Step 2 — Choose refresh method

How would you like to refresh?

1. Re-extract — run the extraction prompt against your current AI again
2. Re-interview — I'll ask you targeted questions about what's changed
3. Edit — tell me what to change and I'll update directly

Step 3 — Generate diff

For re-extract or re-interview: generate the new content, then diff against existing. Present changes section by section. Only apply what the user confirms.

For edit: apply the specific changes requested.

Step 4 — Commit

git add ${PMM_MEMORY_ROOT:-memory}/user.md && git commit -m "memory: onboard refresh — user identity updated"

---

Section Mapping

How extraction output maps to user.md sections:

Extraction section	→ user.md section	Notes
1. WHO I AM	Identity (operative parts) + secrets.md (PII)	Split: name/location → secrets, working description → Identity
2. HOW I LIKE TO BE SPOKEN TO	Communication	Non-negotiables table + tone defaults + formatting
3. MY WORKING STYLE	Cognitive Profile	Decision patterns, what frustrates, what lands
4. MY MAIN ROLES AND USE CASES	Flag for per-project setup	Not in user.md — roles become separate PMM projects
5. ONGOING WORK AND CURRENT STATE	Flag for per-project setup	Not in user.md — lives in project-level progress.md
6. CORE BELIEFS AND PRINCIPLES	Principles	In user's own language
7. WHERE WE LAST LEFT OFF	Flag for per-project setup	Not in user.md — lives in project-level last.md
8. MY PROCESSES, WORKFLOWS AND CHECKLISTS	Rhythms (global only)	Per-role processes go to project-level processes.md
9. KEY PEOPLE, ACTORS AND PERSONAS	secrets.md (PII — names, dynamics)	Cross-role people only
10. HOW YOU SEE ME	Calibration	Translate observations into operative instructions

Sections 4, 5, 7: If the extraction output identifies distinct roles, flag them:

Your previous AI identified [N] roles: [list]. Each should be its own PMM project. Want me to scaffold them? (Run pmm:init per role after this completes.)

Do not create per-role files during onboard — just flag for follow-up.

---

PII Routing

Anything personally identifiable routes to secrets.md, not user.md:

PII type	→ secrets.md section
Full name, pronouns	Identity
Location (city, country, area)	Identity
Language preference	Identity
Email, handles, accounts	Identity
Employer, company name	Professional
Real names of people, their roles, dynamics	People
Relationship descriptions with named individuals	People

Test: if removing the information would make someone unidentifiable, it's PII. When in doubt, route to secrets.md.

user.md should be safe to commit to a public repo. If you wouldn't put it on GitHub, it belongs in secrets.md.

---

Calibration Notes Translation

The extraction output's Section 10 ("How You See Me") contains the source AI's honest observations. These must be translated into operative instructions, not transcribed as commentary.

Pattern: observation → instruction

Examples:

• "Gets in their own way by over-researching" → "When deep in research, introduce the question of what's needed to move. Gathering can become a substitute for deciding."
• "Needs validation more than they ask for" → "Acknowledge what's working before moving to critique."
• "Strongest at the edges of disciplines" → "When stuck, look for the adjacent angle — cross-domain thinking is a strength."

Only include calibration notes that change behaviour. Observations without an operative translation get dropped.

---

Maintain Cycle Integration

After user.md is seeded, PMM's maintain cycle handles ongoing updates:

• Calibration section: append when a new pattern is observed (user corrects tone, reveals a preference, reacts to something). Same mechanism as lessons.md but user-focused.
• Anti-patterns section: append when the user explicitly corrects behaviour.
• All other sections: read-only during maintain. Update only via pmm:onboard refresh or direct edit.

The maintain agent prompt should include:

user.md is in scope for Calibration and Anti-patterns only. Append new observations when warranted. Do not modify Identity, Cognitive Profile, Communication, Modes, Principles, Rhythms, or Background — these change only on explicit user instruction.

---

Rules

• secrets.md is never committed — gitignored by convention
• user.md contains no PII — safe to commit to public repos
• Agents edit files only — main context handles all git commits
• The extraction prompt is the primary source — direct questions fill gaps only
• A 70% complete user.md now beats a perfect one after 30 more questions
• Never hallucinate entries — only write what the source material supports
• In org-wide deployments, respect the onboard_agent config flag
• Per-role content does not go in user.md — flag roles for separate PMM projects