> ## Documentation Index
> Fetch the complete documentation index at: https://docs.profy.cn/llms.txt
> Use this file to discover all available pages before exploring further.
# End-to-end: Daily briefing
> Build a briefing pipeline that runs itself, with the real limits of recurrence and a cost brake
This builds a daily briefing that actually runs on its own: at a fixed time each morning, an agent gathers information in a domain you care about, writes it up, and delivers it somewhere you'll see it.
Scheduled tasks are the only feature on Profy that **spends your money while you're not there**, so this guide is less about how to create one and more about **making it run on the schedule you think it does** and **noticing when it goes wrong**. The recurrence options are far narrower than natural language suggests; knowing that up front avoids most of the rework.
## What you'll get
A new message in your conversation with that expert at 08:30 every day (time configurable), containing that day's briefing. If you create it from an IM channel like Feishu, the briefing is pushed back into that conversation too.
## The key constraint: there are only four recurrence types
This is the most important section in this guide. A scheduled task's repeat type has exactly **four possible values**:
| Repeat type | Meaning | Trigger rule |
| ----------- | ----------- | -------------------------------------------------------- |
| `no-repeat` | One-off | Runs once at the given date and time, then it's done |
| `daily` | Every day | Runs at the given time every day |
| `weekly` | Every week | Runs on **the weekday of the first execution date** |
| `monthly` | Every month | Runs on **the day-of-month of the first execution date** |
Several consequences follow that you need to know:
* **"Mondays and Fridays" is not supported.** One task maps to one weekday. Two days means two tasks.
* **"Weekdays" is not supported.** There is no weekday concept. Either create five tasks, or use `daily` and have the prompt make the agent say "it's the weekend, skipping" — but that still executes and bills every single day.
* **Hourly or minute-level recurrence is not supported.** The tightest legal cadence is once a day. "Every 30 minutes" cannot be done.
* **Be careful with the 31st.** If the first execution is on the 31st, months without a 31st simply won't fire. For "last day of the month," the 28th is the safer choice.
Times are interpreted in `Asia/Shanghai` by default; another timezone can be specified at creation.
## Prerequisites
A scheduled task belongs to an expert. Choose one you'll actually open regularly, because results land in your conversation with it.
**Do not create the scheduled task first.** Run the briefing manually in conversation, confirm the output quality and the cost of one run, and only then freeze it into a schedule. Skip this and you get a task that burns credits on time every day producing output you don't want.
After that manual run, check what the conversation consumed. Monthly cost ≈ per-run cost × runs per day × 30. Make sure you're comfortable with that number before continuing.
## Steps
### Step 1: Run it manually and finalize the prompt
```text theme={null}
Give me today's AI product briefing:
Scope: product and pricing changes at OpenAI / Anthropic / Google DeepMind,
plus equivalent moves from major Chinese vendors
Window: last 24 hours
Structure:
1) The three that matter most (one-line conclusion + link each)
2) Also worth knowing (list, max two lines each)
3) If nothing notable happened today, just say "nothing notable today" —
do not pad with older news
Length: under 500 words
```
Iterate until you're happy, then **save that exact prompt verbatim** — it becomes the scheduled task's prompt.
The "say nothing when there's nothing" constraint is essential. Without it, the briefing degrades on quiet days into a repackaging of three-day-old news, and it can take weeks before you notice.
### Step 2: Create the task
Just say so in conversation:
```text theme={null}
Turn that briefing into a scheduled task:
- Title: AI product daily
- Every day at 08:30 (Asia/Shanghai)
- Use exactly the prompt I confirmed above, unchanged
```
The agent calls the `schedule` tool and returns a card with the task ID, execution time, and repeat type. **Check the repeat type and time on that card** — it's your only chance to confirm it was created the way you meant.
### Step 3: Confirm in the management UI
The agent's `schedule` tool can do exactly two things: **create** and **list**. Editing and deleting must happen in the UI:
| Action | Where |
| --------------------------------------------------------- | ------------------------------------------------ |
| See all tasks, next run time, execution history | The expert's scheduled tasks page |
| Change time / recurrence / prompt / notification channels | Task detail, edit |
| Pause | Task detail, cancel (status becomes `cancelled`) |
| Delete | Task detail, delete |
Don't tell the agent "move the daily to 9am" or "cancel that briefing" — it has no tool that performs those actions. It may politely reply "done," but nothing happened and the task still fires at 08:30. Changes must be made in the UI.
### Step 4: Choose the execution type
There are two task types at creation, and the cost difference is an order of magnitude:
| Type | Behavior | Consumes credits |
| ---------- | -------------------------------------------------------------- | ---------------------------- |
| `execute` | Actually runs a full conversation and produces content | Yes, billed as `token_split` |
| `reminder` | Pushes the text you wrote at the scheduled time, no model call | No conversation credits |
A briefing needs `execute`. If what you actually want is "remind me to go look at it myself," `reminder` costs nothing.
### Step 5: Configure delivery
Results land in your conversation with the expert by default. If you created the task from Feishu / DingTalk / WeCom, the source channel is recorded automatically and the briefing is pushed back to that IM conversation as well.
Tasks created on the web default to in-app only. To push to IM, enable the channel in the task detail.
### Step 6: Check it every day for the first week
For the first seven days, check three things — none of which is content quality:
1. **Did it run at all** — there's a record for that day and its status is `completed`
2. **Did it run at the right time** — actual execution time matches what you set
3. **Is the content new** — compare with yesterday; it should not be the same links reworded
Item 3 is the one people skip and the one that most often goes wrong.
## How execution actually behaves
Things you'll eventually run into, better known in advance:
If a run fails (network issue, insufficient balance, model error), the task is marked failed with a reason and **is not retried**. For recurring tasks the next cycle fires as normal, but that day's briefing simply doesn't exist. So "I subscribed to a daily so I won't miss anything" is not a safe assumption — check the execution history periodically.
If the system is busy at the scheduled moment, the run still happens within the next hour. Misses beyond an hour are skipped entirely.
Scheduled tasks use the model preference saved on your account, falling back to the platform default. Temporarily switching models in some other conversation doesn't affect it.
A run with insufficient balance fails and records the reason, but nothing separately alerts you. Keep a buffer for long-running schedules.
A task can declare a predecessor: it runs only after the predecessor completes. If the predecessor fails you can choose to skip or continue anyway. Waiting for a predecessor is bounded (about an hour), after which it's treated as failed. Useful for "fetch data first, then write the report."
## Boundaries and failure modes
| Symptom | Cause | Fix |
| --------------------------------------------------- | -------------------------------------------------------- | -------------------------------------------------------------------------- |
| Didn't run at the scheduled time | Status isn't `pending`, or the task was cancelled | Check status in the UI; `cancelled` tasks never fire again |
| Weekly task fires on the wrong weekday | The weekday is derived from the first execution date | Change the first execution date to the weekday you want |
| A month was skipped (monthly) | That month has no such day-of-month (e.g. the 31st) | Move the day-of-month to 28 or lower |
| Briefing nearly identical to yesterday | Prompt lacks the "say nothing when there's nothing" rule | Add it and update the prompt in the UI |
| Told the agent to change the time, nothing happened | The agent has no tool to modify tasks | Change it in the UI |
| Several days of failures in a row | Insufficient balance or model unavailable | Read the error in the execution history; top up or change model preference |
| Monthly cost far above expectations | Frequency or per-run cost underestimated | Lower the cadence, narrow the window, cap output length |
## Related pages
Full reference for scheduled tasks
Deliver briefings into Feishu
Get the one-off flow working first
Every hard limit on the platform