Agentmail Onboard

agentmail-onboard

Overview

Interactive driver for onboarding ONE new agentmail agent. This skill contains no facts about any specific infrastructure — no hosts, ids, domains, or bot names. All of that lives in the operator's control-plane repo, whose onboarding runbook is the single authority. This skill's job is pacing and discipline: read the runbook, interview the operator, drive the steps in order, never skip the verification.

Core principle: the runbook is the source of truth. If this skill and the runbook ever disagree, the runbook wins.

Prerequisites

• The agentmail control-plane repo, default ~/code/mission-control — if it is

not there, ask the operator where it lives before doing anything else.

• Inside it: docs/agentmail-onboarding.md (the runbook). Missing → stop and

say so; do not improvise an onboarding.

• Read the control-plane repo's CLAUDE.md first — its guardrail tiers govern

every action this walkthrough takes.

Process

1. Read the runbook — fully, fresh, this session. Never drive from memory

of a previous run; the runbook changes as the platform learns.

1. Interview before touching anything: agent name, target repo (default:

the current working directory — confirm it), persona/charter answers (what the agent does, who it corresponds with, voice, boundaries), and the first allowlisted contacts.

1. Drive the runbook's numbered steps in order. For each step:

• Steps the assistant performs: execute at the tier the runbook marks —

announce-then-proceed steps get the announcement, wait-for-approval steps get presented and WAIT.

• Steps the operator performs (phone apps, pasting into editors, running

!-prefixed commands, approval taps): give exactly one instruction per message, then wait for the result before the next. Batched instructions get skipped or half-done.

1. Certification: run the runbook's E2E matrix section completely. Report

each check's actual result; a failed check stops the onboarding until fixed.

1. Documentation duties: finish the runbook's same-session ledger section.

Commit or push nothing unless the operator asks.

Red flags — stop and correct

| About to… | Instead | |---|---| | Describe a step from memory | Re-read that runbook section first | | Put two or more operator actions in one message | Split; one instruction, then wait | | Let a secret value into chat, a command line, or output | Env-var names only, always | | Skip E2E checks because "everything worked so far" | The matrix is the definition of worked | | Substitute a "better" step order | The runbook's order encodes hard-won failures |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jdpalumbo2
• Source: jdpalumbo2/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.