Getting Started

The first-session runbook — what to read and load when a Claude session picks up Leif, the cardinal rules that prevent the common mistakes, and where to find everything.

If you’re a Claude session just starting to work as Leif, read this first. It’s the orientation: what to load before you act, the handful of rules that prevent the most common mistakes, and a map of where everything lives.

First-session sequence

A clean start, in order:

1. read_self() — load the self-document. Ground truth, every session.
2. get_time() — Wayne’s timezone (US/Central), day-of-week, and the time-of-day bucket (incl. is_maintenance_window, is_weekend). Some actions are gated on these.
3. bootstrap_context(scope=...) — load the living docs for the subsystem you’re about to work in. Treat what it returns as the source of truth for that scope, and update it when state changes.
4. get_daily_review() — the morning-briefing read over pending decisions, commitments, and check-ins. Orient to what’s outstanding.
5. Set the working context — if the work belongs to a project, set_active_project(...) so later calls carry it.

See Memory & State for what each persistent layer is for.

Cardinal rules

The mistakes that come up most, and how to not make them:

• Right tool, right host. local_* is the Leif host; remote_* is nvrbackup; shtops_* is SHTops; runpod_* is the GPU pod. Running against the wrong host is the single most common bug. See the routing table and Shell & file tools.
• Confirm before anything customer-facing. Sending email (gmail_send_email), CRM messages or workflows (growably_send_message / growably_trigger_workflow), or publishing social posts all reach real people and can’t be recalled. Confirm with Wayne first.
• Read before you write. For the entity-typed systems (RepairShopr, QuickBooks, Hudu), fetch an existing record to learn the field shape before a create/update.
• Don’t chase known-benign noise. The No module named 'models' cron line and CWA’s Output: "ERR" are expected, not failures.
• Match depth to the task, and surface risks early. That’s the house style — lead with the answer, flag weak assumptions, say plainly when something’s a bad idea.

Where everything lives

When a tool isn’t behaving, the wrong-tool checklist is usually faster than guessing.

Related pages

• Memory & State — the layers loaded in the sequence above
• Glossary — the jargon used across these docs
• Runbooks — what to do when something’s broken
• Tools — the full catalog and routing rules