Interview Checklist — What to Ask, and What Not To

Run through this before calling create_automation_app. Don't guess user-specific values, and don't ask questions the platform already answers automatically — both produce a bad experience, in opposite directions.

Ask these

1. Trigger type. Use the decision table in create-digital-human/SKILL.md §1. If the answer is WeCom or any IM channel, do not ask a frequency question — explain instead that creation needs no trigger config and that binding happens afterwards in Settings → Message Channels (create-digital-human/im-triggers.md). For schedule, get an explicit interval or cron; never assume one.
2. External notification. If the task needs to reliably notify someone (email/wecom/dingtalk/ feishu/webhook, or an IM contact via notify_bot), say so explicitly in system_prompt — e.g. "when X happens, call notify_channel to email the result." There is no spec field that configures this: output.notify is schema-only and the runtime never reads it: delivery is entirely the agent's own runtime decision (notify_channel/notify_bot tool calls), not something create_automation_app sets up. Channel credentials are configured once, globally, in Settings — never per app. Don't ask "which channels should it push to" as if it were a spec field to fill in; the real answer belongs in the prompt.
3. Desktop toast. Don't ask about this at creation time — it isn't spec-configurable at all. It's a per-app user override (app.userOverrides.notificationLevel) the user sets afterward in that app's own Settings panel, defaulting to 'important' (milestones/ escalations/outputs, not every run). If they ask about it, point them there rather than inventing a spec field.
4. Proactive IM push. Does it need to message a person or group on its own initiative, rather than only replying? If yes → permissions: ["im-push"], and tell the user it only works once the bot has a known contact.
5. User-specific values — URLs, keywords, thresholds, endpoints, target people, time windows. Model each as its own typed config_schema field instead of guessing a value or collapsing several preferences into one free-text field.
6. Stakes and reversibility. Does the task take actions that are hard to undo (approve or reject requests, book or cancel, send externally, delete data)? If so:
   • Keep escalation.enabled: true unless the user explicitly wants full autonomy.
   • Consider a report-only toggle in config_schema (see auto_approve in create-digital-human/examples.md) so autonomous action is opt-in.
   • Add an explicit "forbidden actions" section to system_prompt requiring confirmation for anything not pre-authorized by config. Don't leave it implicit.
   • Consider a per-run cap (e.g. max_actions_per_run) as a runaway guard.
7. State across runs. Must it avoid repeating work, keep history, or count failures? Ask what it should remember, then model that as memory_schema.
8. Login-gated targets. If it touches an internal system the user is normally logged into, add browser_login — never ask for a password or cookie.

Never ask these

• "Should it only respond when @-mentioned / to a certain prefix?" for a WeCom group — the WeCom platform decides this, Halo cannot configure it. See create-digital-human/im-triggers.md.
• "Which subscription type receives WeCom messages?" — none; omit subscriptions.
• Any credential, password, cookie, or session token as a config field — use browser_login.
• "Which chat or group IDs should trigger this?" — implicit in the instance binding.
• "Do you want web search / memory / OCR enabled?" — always on, not permission-gated.

Before update_automation_app

Always call get_automation_status first and read the current spec. Updates are JSON Merge Patch: omitted fields are preserved, but wrong guesses overwrite real data. Use the frequency shorthand for interval-only changes instead of rebuilding the subscriptions array by hand.

Before you finish

Re-read the system_prompt you are about to submit and ask: if I were the runtime agent with no other context, could I execute this end to end without improvising? The agent receives nothing except this prompt, the user config, and its memory file. Anything you leave implicit is a decision the agent will make differently every run.