The onboarding agent

The brain is the onboarding doc

This lesson produces an onboard-employee skill that drafts your next hire's entire first week - and a brain that answers their questions so a manager does not have to. (Despite the lesson's name, this is still rung two of the ladder: a skill that generates a ramp file and a checklist draft. The word agent is earned at the capstone, with the brain gardener.) Most onboarding docs are stale the day they are written, because they are a snapshot of a moving company. The brain does not have that problem: it is the live working memory of the company, and every folder self-describes through its CLAUDE.md header.

So the 2026 onboarding pattern has two halves. An onboard-employee skill runs the mechanical checklist. And the new hire's literal first task is: open Claude Code in the company brain and ask it questions. Anthropic's own report describes new hires ramping by having Claude walk them through unfamiliar codebases - the same mechanic works on an ops repo, because to Claude a company is just another well-organized folder tree.

Build the onboard-employee skill

The skill encodes your day-1/week-1 checklist. It is the new-client skill's twin: every step is checkable, which makes it the easiest workflow in this track to trust.

1. Write your real checklist as plain steps first: accounts to request, docs to read in order, people to meet, first tasks to complete. If you do not have one, draft it from the last hire's actual first week.
2. Create .claude/skills/onboard-employee/SKILL.md with that checklist as the body. Take a name and role as arguments: /onboard-employee "Jamie Lee" "ops associate".
3. Make the skill generate a personal ramp file at meetings/onboarding/<date>_<name>.md: their reading list pulled from playbooks/ and company/, their week-one tasks, and who owns what from org-chart.md.
4. Account-creation steps stay human: the skill drafts the access-request list (email, Slack, CRM, billing tool, with the named approver for each from org-chart.md) but a person clicks the buttons.
5. Run it for the next hire - or dry-run it for the most recent one and compare against what actually happened. The gaps you find are checklist bugs; fix them in the SKILL.md.

Keep org-chart.md current and the skill stays current - it reads ownership at run time instead of hardcoding names. This is the facts-vs-procedures split at work: the chart is a fact in the brain, the checklist is a procedure in the skill, and neither duplicates the other.

Day one: ask-the-brain hour

Block one hour on the new hire's first day. They open Claude Code in the brain repo and ask whatever they would normally ask their manager. Give them a starter list and let them go.

• "Who owns the Acme account and what is its status?"
• "What is our pricing for the standard engagement?"
• "How do we run a client kickoff? Walk me through the playbook."
• "What is the expense policy for travel?"
• "What skills exist in this repo and what does each do?"

Have them log every question the brain answered badly or not at all. That list goes to the brain owner, who fixes files - not the hire's manager, who would otherwise answer the same questions live for the fourth hire in a row.

Measure it, then scale it

Time-to-productive is measurable: days until the hire completes a real task without help. The target this pattern aims at is weeks down to days, because the bottleneck was never ability - it was access to scattered context, and the brain removes it. Write down your current baseline before the next hire so you can show the delta.

Do this now

Sources and further reading

• How Anthropic teams use Claude Code
• Claude Code memory docs (subfolder CLAUDE.md loading)