Lesson 4 of 8 · 6 min read

The four building blocks

Skills, connections, schedules and actions: what each one does, and what it is called in Claude and ChatGPT.

Same blocks, different names

Every assistant that can automate work offers the same four building blocks. A skill holds the instructions. A connection gives the assistant read access to the data. A schedule starts the task without you. An action is what the assistant may do with the result.

The products rename these blocks often. Claude merged its Cowork agent into the main app in September 2026. OpenAI switched off ChatGPT’s agent mode in August 2026 and pointed users to ChatGPT Work. If you learn the blocks, a rename costs you a menu search, not a new course.

This lesson covers each block in turn. For each one, it says what the block does, what Claude and ChatGPT call it, and what to watch for. The steps at the end show where to find each block, and the date someone checked them.

Skills: the instructions

You wrote your skill in Lesson 3. Claude calls these skills. ChatGPT also calls them skills, and packages them for sharing as plugins. Google’s Gemini app calls them skills too, and reads the same SKILL.md format.

A skill is the only block you write yourself. The other three are settings. That is why the skill carries the most important rules, such as "this is a draft" and "use only numbers from the sources". A setting can change in an update. Your written rule stays in your file.

Connections: the data

A connection lets the assistant read a source such as your email, calendar, files or team chat. Claude calls these connectors. OpenAI’s pages call them apps or connectors, depending on the page. Both products offer ready-made connections for common services, and both let you add custom ones.

Connect only what the task reads. Dana’s report needs the shipment log, the supplier notes and the team channel. It does not need Dana’s whole inbox. Each extra connection adds data the assistant can see, and Lesson 6 explains why that matters.

Check your employer’s rules before you connect a work account. Some companies allow only approved assistants to read company data. The rule applies even when the product makes connecting easy.

Some connections can write as well as read. A mail connection may be able to draft, send or delete. Where the product lets you choose, allow reading and drafting only. A connection that can send is an action, and the action rules below apply to it.

Schedules: the trigger

A schedule runs the task at a set time without you asking. Claude offers scheduled tasks on its paid plans. They can run hourly, daily, weekly or on weekdays, and they keep running when your computer is asleep. A task that needs files on your own computer runs only on that computer.

ChatGPT also offers scheduled tasks. The number you can keep depends on your plan, so check OpenAI’s scheduled tasks page for your limit. Both products let you run a scheduled task by hand, which Lesson 5 uses for testing.

Actions: what the assistant may do

An action is anything the assistant does beyond writing text: saving a file, drafting an email, filling a document, or sending something. Claude asks before it takes an action, by default. Its agent can keep working after you close your laptop. ChatGPT Work gathers information across your connected apps and files, and it can keep working on a large task for hours.

In this course, the action stops at a draft. The assistant may save a document or create an email draft. It may not send, delete, pay or change a system other people rely on. Keep the default where the assistant asks before acting, and do not turn it off to save clicks.

How each block fails

Each block has its own way of failing, and users report all of them. A skill fails by not running, as Lesson 3 described. A connection fails when its sign-in expires, so the assistant reads nothing and may not say so. That is why every spec in this course tells the assistant to report a missing source.

A schedule fails when nobody is there to answer a question. One user’s scheduled Claude task stopped at permission prompts that waited up to 73 hours, so it ran 16.9% of the time. Another found that the "Always allow" choice did not carry over to later scheduled runs.

An action fails when the product blocks it in an unattended run. On one ChatGPT forum thread, a user found that all 12 scheduled Gmail sends in a run were stopped by a safety check. Another user on the thread reported that the same kind of send worked in a live chat, and that labeling still worked in scheduled runs.

The pattern is the same across products: a task that works while you watch can fail when you are away. Lessons 5 and 6 show how to test the unattended run itself.

Set them up in order

Build the blocks in a fixed order. Install the skill first. Add the connections second, one at a time, and check that the assistant can read each one. Run the task by hand third, more than once. Add the schedule last, once the drafts are good.

This order means each failure has one cause. If you add everything at once and the draft is wrong, you cannot tell whether the skill, a connection or the schedule caused it. Lesson 5 follows this order step by step.

What changed this year

Two retirements affect anything you built before 2026. OpenAI plans to retire custom GPTs on December 11, 2026 and replace them with plugins. Google is replacing Gems with skills. For business accounts, Gems stop working no sooner than March 1, 2027.

Do not build anything new on a retiring feature. If you already have a custom GPT or a Gem that does a recurring task, move its instructions into a SKILL.md. That file works in Claude, in ChatGPT and in the Gemini app.

Worked example · Larkfield Supply

Dana maps the report to the four blocks

Dana takes the status-report skill from Lesson 3 and decides what each block needs.

  1. Skill: the larkfield-status-report file, installed in Dana’s assistant and tested with a dry run on two past weeks.
  2. Connections: the Operations folder for the shipment log, the supplier updates document, and the team chat channel. Nothing else.
  3. Schedule: every Friday at 1 p.m., which leaves Dana two hours to check the draft before the 3 p.m. leadership meeting.
  4. Actions: Dana considers what the assistant should do with the finished draft.
    What would you do next?Which action should Dana leave turned off?Think it through, then show the answer

    Sending. The assistant saves the draft as a document and stops. Dana reads it, fixes it if needed, and sends it to the leadership team.

Dana writes the map at the bottom of the spec as a comment for later: one skill, three connections, one Friday schedule, and an action that ends at a saved draft.That answers the question you guessed at: Every assistant that automates work offers the same four building blocks. One of them holds the instructions. Guess the other three.

Starter prompts

  • Map my task to the four blocks

    Here is the SKILL.md for a recurring task:

[paste your SKILL.md]

Map it to four building blocks. Skill: anything the instructions still need. Connections: list each source it must read, and nothing more. Schedule: when it should run, and how long I need to check the draft before it is due. Actions: what the assistant may do with the result. The result must stop at a draft, so list any action I should leave turned off. Then list what each block needs from my plan, and what I should check on the vendor’s current pages.

Steps for your assistant

In Claude

Checked on October 15, 2026

  1. Skills: Customize, then Skills.
  2. Connections: Customize, then Connectors. In a conversation, the plus button at the lower left, then Connectors, turns each one on or off.
  3. Schedules: Scheduled, in the left sidebar. Paid plans only.
  4. Actions: by default, Claude asks before it acts. Leave that setting as it is for this course.

Free plans can add one custom connector.

In ChatGPT

Checked on October 15, 2026

  1. Skills: in ChatGPT Work, type @ to find your skills, or @skill-creator to make one.
  2. Connections: OpenAI’s pages call them apps or connectors. Connect each source the task reads.
  3. Schedules: scheduled tasks. Check OpenAI’s scheduled tasks page for the limit on your plan.
  4. Actions: ChatGPT Work works across your connected apps and files. Tell it to stop at a draft.

OpenAI’s help pages could not be read directly when this was checked. These steps follow OpenAI’s skills page and published reports.

When not to use AI

  • Connecting an account that holds data you are not allowed to share with a vendor. Check your employer’s policy first.
  • Building on a feature that is due to retire, such as custom GPTs or Gems.
  • Scheduling a task you have not yet run by hand at least twice.
Recall from memoryName the four building blocks and what each one does.Answer it without looking back, then show the answer

A skill holds the instructions. A connection gives the assistant read access to the data. A schedule starts the task without you. An action is what the assistant may do with the result, and in this course it stops at a draft.

Exercise · your project

Map your task to the four building blocks.

  1. List the exact sources your task reads, and connect only those.
  2. Choose when it runs, leaving enough time to check the draft.
  3. Write down the actions it may take, ending at a draft.
  4. Confirm that your plan covers each block, using the vendor’s current pages.

Sources