Part 3 · Working With AI Tools
Building Your First AI Skill/Agent (No Code)
You have already built one agent in this course: your AIRI Mentor, assembled once from instructions and knowledge, reused every lesson since. Now build the same thing for your own work. Every Monday morning you paste the same specification into your AI tool. And again the week after. You wrote a good prompt once. Now you keep re-typing it.
Why This Matters
A good prompt solves a task once. An agent solves it every time, without you re-typing anything. This is Pillar Five in practice, AI-Assisted Creation and Orchestration, where you stop running tools by hand and start assembling them to run for you. In this lesson you will:
- Build a custom AI agent (a Custom GPT, a Gem, a Claude Project, or a Copilot agent) that handles one recurring task from your own work.
- Assemble it from structured instructions and examples, with no programming.
- Test and refine it until its output is reliable enough to use.
What a no-code agent is
The last lesson had you write a specification before the work started. An agent is where that specification lives on permanently. A no-code agent is a saved configuration of an AI assistant. You write the specification once (the role, the standing instructions, a worked example or two) and the platform remembers it. Every major assistant offers this. It is a Custom GPT in ChatGPT, a Gem in Gemini, a Project in Claude, an agent in Copilot. Same idea, and not a line of code.
- Standing instructions: the specification you would otherwise retype (role, task, format).
- Examples: one or two real input-to-output pairs that show your standard.
- Boundaries: what it should flag or refuse, and where a human must check.
One caution is built into the design. An agent is a tool you start for one narrow task and stop when the task is done, and you govern it like one. And because it runs your instructions every time, a blind spot in those instructions repeats every time, and it can still hallucinate. Build a human checkpoint into the task.
A team lead I worked with turned raw meeting notes into a weekly status update by hand every Friday. She built a Project with standing instructions (summarise into three headings: done, blocked, next) plus two of her past updates as examples. Now she pastes the notes, gets a draft in her format, edits, and sends. The judgement of what counts as blocked stays hers.
- Create your own GPT: help.openai.com
- Build a Gem: support.google.com/gemini
- Claude Projects: support.claude.com
- Copilot agents: learn.microsoft.com/copilot
Key Takeaway
Keep each agent to one narrow task. The moment you ask a single agent to do everything, its output turns unreliable: build a second narrow agent instead.
Check your understanding
Self-check only. Nothing is scored and nothing is recorded; open an answer to compare it with your own.
Lin manages procurement and creates purchase order summaries every Friday. She decides to build her first AI agent for this task. She writes these instructions: 'You are a procurement assistant. When I give you a list of purchase orders, create a summary table with columns: PO Number, Supplier, Amount, Status. Then list any POs over $50,000 that need my approval. Use a professional tone.' She uploads her company's PO template as a reference document. After testing, the agent produces a good table but invents fictional PO numbers instead of using her data. What should she fix?
- Switch to a different AI platform, the one she is using cannot handle tabular data.
- Add a constraint to her instructions: 'Use only the purchase order data I provide. Do not generate or invent any data. If information is missing, flag it rather than filling it in.'
- Upload more reference documents so the agent has enough real data to work with.
- Remove the PO template and write longer instructions instead.
Show the answer
Add a constraint to her instructions: 'Use only the purchase order data I provide. Do not generate or invent any data. If information is missing, flag it rather than filling it in.'
Kumar's first agent works well, so he keeps adding tasks to it: the weekly summary, then a competitor watch, then drafting client emails. A month later the outputs are unreliable and he cannot tell which instruction is causing it. What is the mistake?
- He tested it too little at the start; more early runs would have prevented this.
- He let one agent grow beyond one narrow task. The fix is to split the work into separate agents, each with one job and its own checkpoint.
- He chose the wrong platform; some platforms handle multiple tasks natively.
- Agents degrade over time and need to be rebuilt monthly.
Show the answer
He let one agent grow beyond one narrow task. The fix is to split the work into separate agents, each with one job and its own checkpoint.
Saved in this browser only. Carry progress between devices with the resume link on the course page.