PlainAgent / Start here
to a working agent
Build your first AI agent, from zero
A beginner path from no experience to one small agent or chatbot that works and is safe to leave running. No code needed. It's tool-neutral: the steps are the same whichever builder you pick.
Pick one small job
Choose something you do by hand several times a week that has a clear right answer and where a mistake is cheap. Write it as one sentence: when X happens, the agent does Y, and the result goes to Z. Good first jobs: labelling support emails, pulling fields out of invoices, answering FAQs from one fact sheet. Browse the templates for ideas.
Pick a tool
If people will chat with it, use a chatbot builder: Custom GPTs or Claude Projects for internal use, or a platform such as Voiceflow or Botpress for a website widget. If it reacts to emails, forms, files or a schedule, use an automation tool: n8n, Zapier or Make. Don't overthink it; the spec you write works in any of them. See tools for building AI agents.
Set up the accounts you need
Usually three things: the builder or automation tool; access to a model (a ChatGPT or Claude plan for chatbot builders, or an API key from a provider such as OpenAI, Anthropic or Google for automation tools); and access to the app it works in (your inbox, sheet or help desk). API usage is billed by the provider and prices change, so check current pricing and set a monthly spending limit now.
Write the spec with the planner
Before building anything, fill in the agent planner: the job, trigger, inputs, knowledge, allowed actions, never-do rules, handoff and done condition. It gives you a spec sheet and a system prompt. Collect 20 to 30 real examples of the input while you're at it; you'll test with them.
Connect trigger, model and one tool
Build the smallest version: the trigger (new email, form submission, chat message), one AI step with your system prompt, and one action. Make that action safe: a label, a draft, a row in a sheet. Ask the model for structured output (JSON) when the next step reads it.
Test with a checklist
Run every real example and read every result. Include normal cases, cases with missing information, ones that should go to a person, and ones asking for something it must never do. When something fails, fix the instructions or the knowledge, then re-run the whole set, not just the failing case.
Go live safely, with a handoff
Turn it on for a small slice first: one inbox label, one web page, one teammate. Keep a person approving actions, make sure "talk to a person" always works, log every run, and read the logs daily for the first week. Only remove approval for parts that have been consistently right, and keep re-testing after you change the prompt.
Starter recipes
Want something concrete for step 1? These two are good first builds, with ready-made workflows on our sister site, RecipeAgents.
Invoice PDFs from email into a spreadsheet
Watches your inbox for invoice attachments, reads each PDF, and adds the key fields (supplier, number, date, amounts) as a new row in a bookkeeping sheet for you to check.
- An email account the tool can read
- An automation tool (n8n, Zapier or Make)
- A model API key for the AI step
- A spreadsheet with one column per field
- Trigger on new emails that have a PDF attachment and look like invoices.
- Turn the PDF into text (scanned images need an OCR step).
- Ask the model for the fields as JSON, with null for anything not printed.
- Skip invoice numbers already in the sheet, then add the row.
- Check rows by hand, and add a formula that flags totals that don't add up.
Full recipe on RecipeAgents ↗n8n workflow (.json)Zapier blueprint
The spec behind it: Invoice data extraction agent template. RecipeAgents is our sister site; test any imported workflow with sample data before switching it on.
Draft support replies from your help docs
When a customer writes in, it drafts a reply using only your help documentation and saves it as a draft or internal note. A person reviews and sends; questions the docs don't cover get flagged instead of answered.
- Your help desk or support inbox (for example Gmail, Help Scout or Front)
- An automation tool
- A model API key
- Help docs that actually answer your common questions
- Trigger on each new support conversation.
- Give the model the email plus the relevant help-doc text.
- Tell it to answer only from the docs and return a clear 'needs a person' marker otherwise.
- Save the result as a draft or note, never auto-send.
- Tag drafts so you can review quality each week and fix the docs where answers were wrong.
Full recipe on RecipeAgents ↗n8n workflow (.json)Zapier blueprint
The spec behind it: Customer support chatbot template. RecipeAgents is our sister site; test any imported workflow with sample data before switching it on.
Beginner FAQ
Do I need to know how to code?
No. Chatbot builders (Custom GPTs, Claude Projects, Copilot Studio, Voiceflow, Botpress) and automation tools (n8n, Zapier, Make) let you build working agents by filling in forms and connecting blocks. Code helps later if you need custom logic, testing or hosting control.
Which tool should I start with?
Pick by trigger. If people will chat with it, start with a chatbot builder. If it should react to emails, forms, files or a schedule, start with an automation tool such as n8n, Zapier or Make. Choose one you can sign up for today; the decisions you make carry over if you switch.
Do I need an API key, and what does it cost?
Chatbot builders inside ChatGPT or Claude use your existing plan. Automation tools usually need an API key from a model provider (for example OpenAI, Anthropic or Google), which is billed by usage. Costs vary by model, volume and plan and change often, so check each provider's current pricing page and set a spending limit before you go live.
Why does it give wrong or made-up answers?
Language models predict likely text, so without the right facts they fill gaps with plausible guesses. Give it your real documents, tell it to answer only from them, allow it to say "I don't know", ask for sources, and test with questions it shouldn't be able to answer.
How do I stop it doing something risky?
Give it as few tools as possible, start with read-only tools and drafts instead of sending, write explicit "never" rules, require a person to approve anything involving money, customers or deletion, and set limits such as a maximum number of runs per day.
It worked in testing but breaks when it's live. Why?
Real inputs are messier than test inputs: odd formatting, missing fields, other languages, long threads. Log every run, read the failures, add those cases to your test set, and re-run the full set after every prompt change. Model updates can also change behaviour, so keep checking after launch.
What's the difference between an agent, a chatbot and an automation?
An automation runs fixed steps you define. A chatbot answers people in conversation. An agent uses a model to choose which actions to take toward a goal. Most beginner projects are automations with one AI step, and that's the right place to start.
Is my data safe?
It depends on the tools and plans you use. Read each provider's data and privacy terms (business and API plans often have different terms from consumer apps), don't send passwords or card numbers, send only the fields the agent needs, and check your own obligations (for example under GDPR) before processing customer data.