Skills and brand voice

Skills in the workspace sidebar, or /workspaces/:id/skills. Owner-only, same as AI settings. Per workspace. Readable without a subscription; saved changes shape replies once subscribed.

Tone

Professional, warm, conversational, or direct.

Response length

Concise, balanced, or detailed.

Empathy

Restrained, balanced, or expressive.

Emoji

Never, occasional, or allowed.

Language behavior

Match the visitor, or always use the default language.

Default language

Defaults to English.

Preferred terminology and prohibited phrases

One per line.

Custom brand guidance

Standing notes for this workspace. They cannot override safety.

Always on. You cannot disable them.

Grounded answers

Factual claims come from approved knowledge or a verified Tool result; otherwise the agent says so and offers a next step or a human.

Human handoff

Honors "I want a person," and hands off when the question is unsafe or unresolved. It summarizes; it doesn't pretend a handoff happened unless request_human succeeded.

Safe connected actions

Only enabled Tools, waits for confirmation on actions, and doesn't claim success before a real result. The card shows how many Tools are enabled.

Off until you turn it on. Diagnosis comes from knowledge and verified live data; the agent remembers steps already tried and stops after a bounded number of failures.

Steps appear in the widget as a card, not a paragraph, with That worked and Still broken controls. Each “still broken” counts one attempt on Doorwick's side, the assistant is told what was tried so it never repeats a step, and when the limit is reached Doorwick hands the conversation to a person with the tried steps in the handoff note. The assistant cannot miscount; the limit lives in the server.

Step style

One step at a time shows a single step per card. Checklist shows up to six steps the visitor can tick off before reporting.

Maximum attempts

1–10, default 3, then hand off with a short summary.