GPTs, agents and skills · AI Tools Field Guide

Generative or agentic?

Two words you will hear. They describe how much of the work the tool does between your messages.

Generative: it writesOne request, one answer, and it stops. You read the draft, decide what is wrong and tell it what to fix. Every plain chat works this way, and it covers most staff work.
Agentic: it actsYou give a goal. It plans, searches, reads what it found, picks the next step and repeats until it is done, using tools on the way. Deep Research, Gemini agents with connectors, Grok skills and automations work this way.

The model underneath is the same; the tools and the loop are what is new. What changes for you: give a goal and a stopping point, expect minutes instead of seconds, and read the steps as well as the result. None of these tools sends email or signs anything on your behalf by default; you still release the product.

Three names, one idea

Agent — Gemini's Agent DesignerA name, a description, instructions, a model, connectors and knowledge files. Shareable with people in your organization; can be put on a schedule (preview).
Custom GPT — ChatGPTSaved instructions plus reference files, opened like a model of its own. Good rule of thumb: a project for ongoing work, a GPT for a reusable assistant.
Skill — GrokA reusable playbook of instructions, attached to any chat from the plus menu: a specialty SOP, not a full workspace.

Every one of them is the same four parts: instructions (what it does, how it behaves, what it must never do), files it may read, a model, and tools it may use. Each tool lays those parts out differently.

The agent's Details panel in Gemini Enterprise's builder: Name, Description, Instructions beginning with a Role paragraph, a Model selector showing 3.5 Flash, and a Connectors section, beside the Flow canvas. — A Gemini agent's Details panel, generated from the starter "military award writing agent".

ChatGPT's GPT editor: Name, Description, Instructions, Conversation starters, Knowledge with an Upload files button (up to 20 files at a time, 40 total, 50 MB each), Recommended model, and Capabilities, with a Preview pane on the right and a Create button. — The ChatGPT GPT editor, empty.

A worked example: Instruction to Checklist

The easiest GPT to explain is one that does a single job everybody needs. This one takes any instruction and returns a numbered checklist with the paragraph reference beside each item. It was built in ten minutes.

The GPT editor filled in for a GPT named Instruction to Checklist: a description, four short paragraphs of instructions, and a conversation starter. In the Preview pane on the right a sample instruction has been pasted and the GPT replies with a six-item numbered checklist, each item ending in its paragraph reference, then a Deadlines found line and an Assumptions line. — The editor filled in, and a test in the Preview pane: a sample instruction goes in, six checklist items with paragraph references come out, then the deadlines it found and the one assumption it made.

The instructions are four lines: what it is ("you turn instructions and directives into action checklists"), what to produce (a numbered checklist, one action per line, the paragraph in brackets), the rules (use only the text provided; never add a step, date or name; write [CONFIRM] for gaps; under 20 words per item) and a closing summary (deadlines found, assumptions). The full text, box by box, is in the Intermediate lesson Build a GPT and a project.

The finished GPT, Instruction to Checklist, opened like a chat: its name and description at the centre, a conversation-starter button reading Turn this instruction into a checklist, and the usual chat box with the model selector GPT 5.6 Terra. — After Create: the GPT opens like a chat, with its starter button.

When to build one

Build a helper when…	Stay with a chat or a project when…
Several people do the same task	Only you do it
The instructions are long and stable	The job changes every time
The output must look the same every time	A one-off draft is enough
New people join and need the same quality on day one	The rules live in your head and you are the only one who needs them

What a helper cannot do

A helper can	A helper cannot
Run its instructions when a person asks, or on a schedule (Gemini agents, Grok automations)	Read your email or your share drive; it sees only what is attached or connected
Work from the files you gave it and from its connectors	Verify its own output; the reader still checks
Be shared with named people (Gemini agents today; a custom GPT's menus offer only Pin and Edit, and Grok project invites were not working in September 2026)	Be shared by a public link
Produce text, tables, and in Grok a file	Remember last week's conversation

An agent does what its instructions say and remembers nothing between visits: a person supplies the inputs, the agent drafts, a person reviews. Google's public name for this in Gemini is Gems: the same idea of a saved role you build once. Building each kind, step by step and with the screens, is in the Intermediate path: Build an agent, Build a GPT and a project, and Grok skills and automations.

Vendor facts checked 25 Sep 2026; re-verify before relying on them.