Many of the routines a team relies on started as a question someone kept asking. What needs attention before the Monday call? Which orders are at risk this week?
At first someone pulls the numbers by hand. Then they ask an AI assistant. Eventually it becomes clear that the question has a fixed shape and could answer itself.
The schedule is the easy part. A routine that runs without anyone watching has to get the definitions right every time, and say plainly when something is missing.
Turning the question into a routine is where most attempts stall. This article follows one invented example, an operations lead at a packaging distributor with two warehouses, from a chat question to a routine a whole team uses. Every step uses the same Hub, the shared foundation of instructions, reference notes and tools behind every Hubzoid agent.
- Question in chat
- Shared definitions
- Weekly routine
- Run record
- Her own assistant
- Team correction
Keep it a question until the answers are right
Every Monday, the operations lead asks the Hub's web chat what needs her attention before the operations call. A useful answer is short:
- late shipments against the threshold
- orders at risk this week
- anything the weekend shift left open
- a plain note about any input that was missing
She keeps asking in chat for a few weeks before automating anything. Each time the answer is wrong, vague or needs a follow-up question, she notes why. Those notes are the most valuable thing this stage produces, because they show exactly what the agent does not know yet.
Most of the gaps turn out not to be about data. The numbers come from tools that read the warehouse system. The gaps are definitions: what counts as late, which orders are at risk, what the weekend shift is expected to hand over.
Write the definitions down once, where every answer reads them
The fix is a short reference note in the Hub that defines each term and the point at which it gets flagged. For the distributor, the invented definitions look like this.
| Term | What it means | Flag it when |
|---|---|---|
| Late shipment | Dispatched after the promised date | More than 4 percent of last week's shipments |
| Order at risk | Due within 48 hours, stock not yet picked | Any order above 10,000 |
| Open handover | A weekend shift item with no named owner | Always |
The note names its owner, the operations lead, and the date she last reviewed it. It also carries two rules that matter more than they look:
- Every figure names where it came from and when.
- If a source fails, the briefing says so and leaves that section empty instead of guessing.
The chat answer, the scheduled briefing and any assistant connected to the Hub all read the same note, so they agree on what "late" means.
A routine is the same question on a schedule
Once the chat answer is reliably useful, it should arrive without anyone asking. In Hubzoid, the simplest way is a scheduled task written in plain language: when to run, where it may save its output, and a few lines of instructions for the Hub's own agent. It uses the same instructions, notes and tools as chat. Nothing is copied.
The lead's task runs early every Monday and saves the briefing to one agreed place in the Hub. A few safeguards come with it:
- No duplicates. It keeps a small record of which weeks it has already written, so a rerun does not produce a second copy.
- No hanging. Each run has limits on how long each round of work may take and how many rounds it gets, so a confused run ends as incomplete.
- One catch-up. If the server was down at the scheduled time, the task runs once to catch up when it starts again.
Some routines need exact steps, for example a threshold that must be calculated before the model writes a word. Those can become a coded workflow, where finished steps are saved so a restart does not repeat them. Starting with the plain-language version, and moving only the parts that need exactness, keeps the routine readable by the person who owns it.
Every run leaves a record someone can read
A routine nobody can inspect is hard to trust. Each run appears in the Console's run history with its status, timing, result and steps, and a failed run shows its error in the same place. When the operations lead wonders whether this week's briefing ran, she can look instead of guessing.
That record also keeps conversations about the routine factual. If the briefing did not arrive, the Console shows whether the run failed, ran long or missed its slot while the server was down.
A finding can be followed up in her own assistant
One Monday the briefing flags late shipments above the threshold at the north warehouse. The lead wants to dig in with the AI assistant she already uses for analysis, such as Claude Code or Codex. A Hub can serve its notes and tools to assistants like these over MCP, the open protocol many assistants use to connect to outside tools. She creates a personal key in the Hub's web chat and connects her assistant.
Her assistant brings its own model, and the conversation stays in her assistant. Every call it makes to the Hub runs as her, under her own access. Most importantly, it reads the same definitions note. She is not pasting a copy of the reporting rules into another tool, where it would quietly drift.
A teammate's correction improves every answer
The warehouse supervisor owns the next decision. He asks the agent in Slack what this week's briefing flagged for the north warehouse. The agent can read the briefing the routine saved, so he gets the answer where he already works.
He also notices something the definitions missed. One of the late shipments was held at the customer's request, and the note does not exclude customer holds.
The fix is not a better prompt in one conversation. It is a change to the shared note that every answer reads.
Reviewed like any other change to the Hub, the correction reaches next Monday's routine, the chat answer and the lead's assistant at once. To keep the rule from slipping back, the team adds a written check: a question about a shipment held at the customer's request, and the expected answer, that it does not count as late. Hubzoid calls these evals. The team runs the check after each change to the definitions, so a later edit that breaks the rule shows up before the next briefing does.
The routine is only as good as the definitions behind it
What turns a weekly question into a routine a team relies on is the work around the schedule:
- asking in chat long enough to find the gaps
- writing the definitions once, where every answer reads them
- keeping a record of each run
- sending corrections back to the shared note instead of into one conversation
Workflows and schedules explains how scheduled tasks and coded workflows run. Connect an assistant covers using a Hub from Claude Code, Codex and other assistants.