# How to Build an AI Agent With ChatGPT or Claude, Step by Step > Build a first AI agent with no code: a one-page brief, a ChatGPT or Claude Project, a ten-request test, and the point where a dedicated builder helps. _Source: https://professionalstoolkit.com/articles/how-to-build-an-ai-agent — The Professional's Toolkit · updated November 2026_ _Part of: [Build a first AI agent that does one job from start to finish](https://professionalstoolkit.com/solutions/build-your-first-ai-agent) — the guide, the do-it-yourself version and the tools that do it_ --- An AI agent is an AI model given four things, instructions, knowledge, tools and a trigger, so it can carry one recurring task from start to finish. Here is how to build an AI agent without code: write the job down as a one-page brief, load that brief and your reference files into a ChatGPT or Claude Project, then test it on ten real requests before you rely on it. The build takes an afternoon. What you get is a capable assistant that waits for you to open it and type, and that is exactly where its limits begin. **10** — real past requests to run through a first agent before you rely on it - **4 parts** — instructions, knowledge, tools and a trigger - **5** — projects a free Claude account can create - **Dec 11, 2026** — standard retirement date OpenAI lists for custom GPTs ## How to build an AI agent: start with the job, not the tool Most first agents fail for a dull reason. Nobody defined the job. Pick one task that comes back at least weekly, has a clear input and a clear output, and where a wrong answer gets caught before it does harm. Drafting replies to quote requests is a good first job. So is turning meeting notes into a follow-up email, or sorting new leads against your own criteria. Anything that sends, books, pays or publishes with no person reading it first is a bad one. If the term itself is still fuzzy, read [what an AI agent is](/articles/what-is-an-ai-agent) first. Then write the brief. One page is enough: 1. **Role and goal.** Who the agent is working for and what a finished job looks like. 2. **Input.** What you will hand it: an email, a form entry, a transcript. 3. **Steps.** Written as *when this happens, do that*, in order. 4. **Output.** The exact format: subject line, length, sections, tone. 5. **Boundaries.** What it must ask about instead of guessing, and what it must never promise. 6. **Examples.** One reply you would send as written and one you would not. That structure follows OpenAI's own guidance for writing instructions: explicit steps, concrete things to do rather than long lists of prohibitions, and brief examples of acceptable and unacceptable output.[1] The same page draws a line worth keeping. Rules, tone and workflow belong in the instructions. Reference material belongs in files. > 💡 **Keep rules and facts apart:** Instructions say how the agent behaves; files hold what it looks up. When a policy changes, replace the file and leave the brief alone. ## The ChatGPT route: a Project, now that the custom GPT is winding down For two years the standard answer was a custom GPT. Check before you plan around one. OpenAI's help page now says the company plans to retire custom GPTs and move creators to plugins, lists December 11, 2026 as the standard retirement date, and states that personal accounts, from Free to Pro, cannot create or publish new GPTs. Existing ones stay usable until their retirement date. For a new build today, the place in ChatGPT that holds instructions and files together is a Project.[2] 1. Select **New project** in the sidebar and name it after the job. 2. Add your files: PDFs, spreadsheets, docs, images or pasted text. Start with your service list, your policies and three past replies you were happy with. 3. Open the more options menu (•••), select **Project settings** and paste your brief as the project instructions. They apply only inside this project and override your global custom instructions. 4. Optional: select **Add sources** and paste a Google Drive file or folder link, or a Slack channel link, so the project can read it as context. 5. Start a chat inside the project and paste a real request. When a response is worth keeping, such as a decision note or a model reply, open its message menu and select **Add to project sources**. How many files a project accepts depends on your plan, so keep the set small and current rather than uploading everything you own. ## The Claude route: instructions plus a project knowledge base Claude's version is built on the same two parts, and it is open to free accounts, which can create up to five projects.[3] For anyone searching for a free AI agent to learn on, that is a fair place to start. 1. Click **Projects** in the left panel, then **+ New Project**. 2. Give it a name and a description. Claude does not see these two fields, so do not hide instructions in them. 3. In the project knowledge area, click **+** and upload your documents, text files or code snippets. 4. Click **Set project instructions**, paste your brief and click **Save instructions**. 5. Start a chat inside the project. Plan around one behavior. Anthropic's page notes that context is not shared across chats within a project unless the information sits in the project knowledge base. So when a chat produces a rule you want to keep, move it into the instructions or the knowledge, not into your memory of that conversation. Gemini users have the same pattern under another name: Google describes Gems as customized versions of Gemini for repetitive tasks, and the same one-page brief works there. > 💡 **Save every request that failed:** Each time a real request trips the agent, add it to your test set. The ten grow into a regression check you rerun after every edit to the brief. ## Ten real requests decide whether it works Take a two-person home-cleaning company. The job: turn quote-request emails into draft replies. The files: a service list, the ZIP codes the crew covers and the cancellation policy. Now pull ten real emails from last month, and choose the awkward ones on purpose. One from outside the service area. One asking for a service you do not offer. One too vague to quote. One angry. One with two questions in a single line. Run each in a fresh chat. Score every draft on three points: the facts match your files, the format matches the brief, and the agent asked a question where the email left a gap. Then fix the brief, not the chat. A correction typed into a conversation is not something to count on in the next one, while a line added to the instructions applies to every run. OpenAI gives the same advice for this stage: before adding more tools, tighten the instructions and add examples. Rerun all ten after each change. A reasonable bar: eight drafts you would send after light edits, and two where it asked instead of inventing an answer. Agents make mistakes. In this setup the human check is simple, because you read every draft and you press send. > ⚠ **Do not let version one send:** Keep the first agent drafting only. You read, you edit and you send, until its misses on your test set are rare and harmless. ## Where an agent that lives in a chat window stops Count the four parts again. You have instructions and knowledge. Tools are limited to what your assistant connects to, and the trigger is you. That works for one person handling a few requests a day. It strains when requests arrive faster than you can paste them, when the result has to be written into a calendar or a CRM, when a colleague needs to run the same agent, or when you need to see what it did last Tuesday. ChatGPT closes part of the gap by itself. Scheduled tasks run one-time or recurring work, and event-triggered tasks can respond to Gmail, Slack or GitHub activity on Plus, Pro, Business, Enterprise and Edu plans, though not on Free or Go.[4] Read the fine print on that page, though. A task created in a project cannot access the files uploaded to or stored in that project. The trigger and the knowledge do not meet. That is the threshold in plain terms. Once you want the agent to start without you, read your own files, act in your apps and wait for your approval on the risky step, all in one place, you have outgrown the chat window. Nothing is wasted. The brief and the ten test requests move over unchanged. ## Three builders that take the same brief further Dedicated builders add the trigger, the app connections and the approval step. Our [AI agent builder guide](/articles/ai-agent-builder) covers the wider field; the three below start in clearly different ways. ### Lindy — Built by chatting, approvals on by default You describe the task in plain English in Slack, the web app, a text or an email. A skill is described once and Lindy picks it when a request needs it; routines run it on a schedule or an event. Anything with outside impact, such as sending an email, waits for your approval by default. - Where it stops: The free start lasts 7 days, then a paid plan is needed. Lindy's own docs say it can make mistakes and to review actions. - Plan: Plus, Pro or Max - [Set up and teach Lindy](https://professionalstoolkit.com/articles/lindy-ai-playbook) · [Lindy](https://professionalstoolkit.com/tools/lindy) Lindy suits the owner or manager whose work already runs through email, calendar and Slack and who never wants to see a canvas. Your brief turns into the description of a skill, and your ten requests become the first messages you send. Our playbook shows how Lindy [is set up and taught](/articles/lindy-ai-playbook). ### Gumloop — Start from a ready-made gallery agent Pick a preconfigured agent from the use-case gallery and authorize each tool with your own sign-in, or ask the built-in Gumball agent to create one from your description. Triggers then run it without you opening a chat. - Where it stops: No free plan, only a one-time trial of Pro that needs a card. The agent can edit its own instructions by default, with no version history. - Plan: Pro - [From trial to team agent](https://professionalstoolkit.com/articles/how-to-use-gumloop) · [Gumloop](https://professionalstoolkit.com/tools/gumloop) Gumloop fits an operations or marketing person who would rather adapt a working example than face a blank page. One setting deserves attention before the first run: its approval presets are chosen per connector and the default is Always allow, so switch anything that writes to an asking preset. The [walkthrough from trial to team agent](/articles/how-to-use-gumloop) covers the order. ### Relevance AI — Drafts the agent, a person publishes Invent builds the agent, its tools and its triggers from a plain-language description and keeps them in draft until a person publishes. Each tool gets its own approval mode: Auto Run, Approval Required or Let Agent Decide. - Where it stops: No free plan, and the builder runs only in a desktop browser. Cloning a Marketplace template does not copy its knowledge, triggers or secrets. - Plan: Pro or Team - [First agent, step by step](https://professionalstoolkit.com/articles/relevance-ai-tutorial) · [Relevance AI](https://professionalstoolkit.com/tools/relevance-ai) Relevance AI is for a team that wants to inspect the whole build before anything runs, and that expects the agent to grow. Its documentation is frank that agents decide for themselves and may not follow the exact path you expect, which is the reason to keep approvals on. The [first-agent tutorial](/articles/relevance-ai-tutorial) goes step by step. When one agent turns into several that hand work to each other, you are designing [agentic workflows](/articles/agentic-workflows). ## FAQ **Can I still make a custom GPT?** OpenAI's help page says personal accounts, Free through Pro, cannot create or publish new GPTs, and that existing unmigrated GPTs stay usable and editable until their retirement date. Migration uses the latest published version, so drafts and unpublished edits do not transfer. **Is there a free AI agent I can build myself?** A Claude Project is available on free accounts, up to five projects, and gives you instructions plus a knowledge base. What a free account will not give you is an event trigger: OpenAI's help page says Free and Go accounts cannot create event-triggered tasks in ChatGPT, and a free recurring scheduled task runs no more than once per day. **Do I need to code to create an AI agent?** No. A Project needs only a written brief and your files, and builders such as Lindy, Gumloop and Relevance AI are set up by description. Code is the developer lane: CrewAI's open-source framework, for example, needs Python skills. **How long does it take to make an AI agent?** Writing the brief and setting up a Project fits in an afternoon. The real time goes into testing: plan for several rounds of running your ten requests and tightening the instructions before the drafts are reliably usable. ## Two weeks by hand, then the builder question answers itself Build the Project this week from the one-page brief and run it by hand for two weeks. Keep a tally on paper: how many requests came in, how many drafts needed real edits, and how often you wished it had started by itself or written the result somewhere else. If the count stays at a few a day and you read every draft anyway, stay where you are. If the tally fills up with missed triggers and copy-paste, take the brief and the ten requests to a builder, and leave the approval step switched on. ## Sources [1] OpenAI Help Center — https://help.openai.com/en/articles/8554397-creating-a-gpt (Nov 2026) [2] OpenAI Help Center — https://help.openai.com/en/articles/10169521-using-projects-in-chatgpt (Nov 2026) [3] Claude Help Center — https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects (Nov 2026) [4] OpenAI Help Center — https://help.openai.com/en/articles/10291617-scheduled-tasks-in-chatgpt (Nov 2026)