Scheduled tasks and automations

The official reference lives at help.openai.com. For the broader agent feature set this slots into, see openai.com.

Setting up the Monday news brief

Let's build the real thing. Goal: every Monday at 8:00 AM, get a tight summary of the past week's developments in your industry, scannable in 60 seconds.

Step 1: Write the prompt as a standalone job

Since there's no follow-up, the prompt carries all the structure. Open a normal chat and type something like this:

> Every Monday at 8am, search the web for the most significant developments from the past 7 days in enterprise AI tooling. Return exactly 5 bullets, each one sentence, ordered by importance. After each bullet, add a source link in parentheses. End with one line: "Worth watching:" and a single forward-looking item. If nothing notable happened, say so plainly instead of padding.

Notice what's doing the work:

• A real trigger phrase ("Every Monday at 8am"). ChatGPT detects the scheduling intent and offers to create a Task.
• An explicit tool instruction ("search the web"). Tasks can use ChatGPT's web search, but only if the prompt makes clear it should. Don't assume.
• A fixed output contract (5 bullets, one sentence each, sources, one closing line). Recurring output you'll skim should look identical every week.
• A failure instruction. Telling it what to do on a slow news week prevents hallucinated filler.

Step 2: Confirm and refine the schedule

When ChatGPT proposes the Task, it shows the parsed schedule. Adjust the time and recurrence if it guessed wrong. You can also create or edit Tasks from your profile menu under the Tasks view, where every active and paused Task is listed.

If you want to be precise about cadence in your prompt, plain English works ("the first Monday of each month", "every weekday at 7am"). ChatGPT translates it; you verify it.

Step 3: Scope it inside a Project (optional but smart)

If you run this brief for a specific client or domain, create the Task from within a Project so it inherits that Project's custom instructions and files. A Task born inside a Project titled "Acme competitive intel" can be told to weight Acme's competitors automatically, no repetition needed.

Reviewing and tuning the output

The first run is a draft, not a deliverable. Treat the first two or three Mondays as calibration.

What to check:

• Did it use the tool you expected? If the brief reads like it came from training data, not this week, the web search didn't fire. Make the search instruction more forceful and add a date anchor ("only items from the last 7 days").
• Is the format stable? Drift in length or structure means your contract is too loose. Tighten it ("exactly 5 bullets", not "a few bullets").
• Are the links real? Spot-check them. If sources are weak, instruct it to prefer primary sources and to skip any item it can't link.

To edit, open the Task and revise the prompt directly. You don't recreate it. You can also pause a Task (useful during a holiday) and resume later.

A clean tuning loop:

1. Read Monday's output.

2. Note the single biggest flaw.

3. Edit the prompt to fix exactly that one thing.

4. Wait for next run, repeat.

Resist rewriting the whole prompt at once. You won't know which change helped.

Making Tasks more powerful: Connectors and the agent

A web-search brief is the easy case. Tasks get genuinely useful when they touch *your* data.

Connectors

Connectors link ChatGPT to external systems (Google Drive, Gmail, GitHub, Outlook, and others, depending on availability and plan). A Task can read from a connected source at run time. For example:

> Every Friday at 4pm, check my connected Google Drive folder "Weekly reports", summarize any document modified this week into 3 bullets each, and flag anything mentioning "delay" or "blocked".

Now the Monday brief or Friday digest is grounded in your actual files, not just the public web. Authorize Connectors once at the account level; the Task uses them on every run.

The ChatGPT agent

For multi-step jobs, the ChatGPT agent can browse, click, fill forms, and operate a virtual computer to complete a task end to end. A scheduled agent run can do things a single prompt can't, like navigating a site that has no clean search.

The tradeoff: agent runs are slower, can hit login walls or CAPTCHAs, and may pause for your confirmation on sensitive actions. For an unattended scheduled job that needs zero human input, a focused prompt with web search is usually more reliable than a full agent run. ReachReachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → for the agent when the task genuinely requires interaction, not just retrieval.

When to leave ChatGPT and use the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → instead

Scheduled Tasks are excellent for *personal*, *human-readable* automations delivered to *you*. They hit a ceiling fast when you need any of the following:

• Delivery somewhere other than the ChatGPT app (post to Slack, write to a database, trigger another system).
• Strict, machine-readable output that downstream code can parse every time.
• Volume, many runs or many users, beyond personal limits.
• Version control and testing of the prompt itself.

At that point you move the schedule out of ChatGPT and into your own infrastructure (a cron job, a serverless function, a workflow tool), and call the OpenAI API on each tick. You provide the scheduling; the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → provides the intelligence.

Here's the Monday brief as a real APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → job using the Responses API with the built-in web search tool and structured outputs, so the result is guaranteed-shape JSON your code can route anywhere:

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

schema = {
    "type": "object",
    "properties": {
        "items": {
            "type": "array",
            "items": {
                "type": "object",
                "properties": {
                    "headline": {"type": "string"},
                    "source_url": {"type": "string"},
                },
                "required": ["headline", "source_url"],
                "additionalProperties": False,
            },
        },
        "worth_watching": {"type": "string"},
    },
    "required": ["items", "worth_watching"],
    "additionalProperties": False,
}

response = client.responses.create(
    model="gpt-4.1",
    tools=[{"type": "web_search"}],
    input=(
        "Find the 5 most significant developments from the past 7 days "
        "in enterprise AI tooling, ordered by importance. Include a source "
        "URL for each. Add one forward-looking item under worth_watching."
    ),
    text={
        "format": {
            "type": "json_schema",
            "name": "weekly_brief",
            "schema": schema,
            "strict": True,
        }
    },
)

print(response.output_text)

Wrap that in your scheduler of choice and pipepipeAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.View full definition → response.output_text to Slack or email. The strict: True schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.View full definition → means you never get a malformed week. For the full reference, see the OpenAI API docs.

If your automation grows into multiple coordinating steps (fetch, summarize, fact-check, format, route), that's the signal to look at the Agents SDK, which is built for orchestrating tool-using agents in code rather than cramming everything into one prompt.

A quick decision rule

• Just for you, readable, low volume? ChatGPT Scheduled Task.
• Needs your private files, still readable, still for you? Scheduled Task plus Connectors.
• Needs interaction with a website? Scheduled run of the ChatGPT agent.
• Feeds other software, strict format, or runs at scale? Your own scheduler plus the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → with structured outputs.
• Multi-step, multi-tool pipeline? Agents SDK.

Pick the lowest-effort option on this list that actually meets the need. Most people overshoot to the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → when a Task would have done the job in two minutes.

Key Takeaways

• Write Tasks as complete, standalone prompts. No human is there to answer questions, so specify the tool to use, the exact output format, and what to do on a quiet week.
• Calibrate over the first few runs, fixing one flaw at a time, and verify that web search actually fired and that links are real.
• Use Connectors and Projects to ground recurring briefs in your own data instead of only the public web.
• Graduate to the API with structured outputs the moment you need machine-readable results, delivery outside the app, or scale, and reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → for the Agents SDK for true multi-step pipelines.