AI Central

How to Build AI Agents Without Previous Knowledge

Download
AI Central
Review AI summary

An agent is three parts, not a new kind of software. AI Central's guide to building AI agents without previous knowledge reduces every agent to a model, a set of tools and written instructions, then shows how to put all three together inside ChatGPT, Claude or Gemini with no code. The step most people skip, testing on ten real cases before trusting the result, is the one the guide calls your job.

Why a chatbot is not an agent

The guide opens with a failed request. Asked to find three free slots next week and draft an invite, a plain chat replies that it cannot see the calendar and offers a template instead. That is the gap the guide sets out to close: a chatbot answers, an agent does the work. The rest is a seven-step route from one to the other, with a map that sends each reader to the step that fits their level instead of step one.

Brain, hands and a playbook: that's an agent

The three building blocks

Step one names the parts and gives each a single job. The model is the brain that reads the request, plans and decides the next move. Tools are the hands and eyes: web search, your files, email, calendar. Instructions are the playbook: who the agent is, what to do and when to stop. A fourth line sits under the three, and it belongs to the reader: test it before you trust it.

Model first, then tools

The model advice runs against the instinct to save money early. Build with the most capable model you have, note how it does on real tasks, then try a faster one and keep it only if results hold. The guide maps this to each app: a higher thinking level in the ChatGPT model menu, Opus on Claude's paid plans and Sonnet on Free, the most capable model in the Gemini menu. Sorting and short replies are the exception, where a fast model is fine from the start.

Tools get the opposite treatment. The guide splits them into data tools that read files or look up a record, action tools that draft an email or book a slot, and other agents the main one can call on. Its rule is narrow on purpose: only the tools the task needs, because each tool is a permission.

Instructions: role, steps, limits

Step four is the core of the guide. "Be helpful" is not a playbook, so every instruction set gets three parts. Role says who the agent is and who it serves. Steps say what to do, in order, for each case. Limits say when to stop, ask or hand over. The worked example is a furniture shop inbox: sort each email into sales, support or other, suggest products from the catalogue for sales, check the order sheet for support, never promise refunds or dates, and write "Needs a human" when unsure.

The limits line carries the most weight. It gives the agent a defined exit for any case its rules do not cover, and that exit is exactly what the test step checks later.

Building it in the app you already use

Steps five and six move the playbook into a real setup. In ChatGPT it becomes a project with instructions and files; in Claude a project or a skill; in Gemini a skill you run with a slash. The sample setup names the agent, pastes the step four playbook, attaches the catalogue, order sheet and returns policy, and switches Gmail on and web search off. For multi-step jobs across apps, the guide points to ChatGPT Work and to Cowork on Claude's paid plans, and to scheduled tasks in all three apps. Its scheduled example reads the shop inbox every weekday at 8 am, drafts every reply and leaves each draft unsent.

Ten test cases before trust

Step seven is a short loop: collect ten real cases with the right answer, run all ten, mark each pass or fail, fix the instructions, run all ten again. The sample test is a customer whose sofa arrived torn and who wants a refund today. The agent passes because it hands the case to a human, notes the damage, promises nothing and finds the order.

No test, no trust

What to watch out for

  • Keep a human on send. The agent reads, searches, sorts and drafts; you approve anything that sends, pays, publishes or deletes.
  • One wrong rule repeats on every run, so review a sample of the agent's work every week.
  • Tell the agent to stop and ask when a case is unclear, instead of guessing.
  • Give a tool only when the task needs it, since every tool is a permission.

When the chat app is not enough

The guide closes with the next tier: Zapier, Make or n8n to run an AI step on every new email, form or row; the OpenAI Agents SDK or the Claude Agent SDK for your own app; MCP, the open standard for plugging tools into agents, to connect your own tools. The building blocks do not change. It is still model, tools and instructions, plus the same ten test cases. Start with one job you repeat every week, write its role, steps and limits, and test it on ten real cases.