AI for Intermediates: 8 skills to hand AI a whole job

How it works

A chat waits for your next message. An agent keeps going: it decides what to do, uses a tool, looks at the result and decides again. Anthropic's guide to building agents describes them as systems where the AI directs its own process and its use of tools.

That freedom is why the finish line matters. Agents usually stop when the task is done, and Anthropic notes it's common to add a stopping rule as well, such as a maximum number of steps. Zapier's new AI step does a version of this for you: you can require approval before a tool runs, and it pauses for a person if one run goes past 75 tasks.

Calling something an agent in your prompt doesn't make it one. What makes it an agent is the tools it can use and the freedom to pick its next step. Anthropic's advice is to start with the simplest setup that works, which is often one good prompt, and add more only when you need it.

Practice

Write a job card for the bakery assistant. Paste it into Claude, ChatGPT or Gemini with the FAQ and the customer's message.

Job: Draft a reply to each new cake enquiry.
Goal: A draft reply and a two-line summary for staff.
You may use: the cake FAQ I give you. Nothing else.
You may not: send messages, book orders or promise a cake.
Finish line: Stop when the draft and summary are written. Then wait for me.

You've got it whenIt writes the draft and the summary, stops, and sends nothing.

Read

• Building effective agents Anthropic · about 12 min Where "start simple" and stopping rules come from, and the line between workflows and agents. Written for developers; the ideas hold.
• What are AI agents? IBM · about 18 min A plain definition, the main types and everyday examples.
• Zapier Agents is now AI by Zapier Zapier · about 4 min Approvals and a pause for a person, in a tool you can try.

How it works

A skill is a folder with one file called SKILL.md, plus any examples or templates the job needs. The app reads only each skill's name and description until a task matches, then loads the full instructions. That's how it can keep many skills ready without filling its working space.

SKILL.md is an open format. Anthropic started it and released it as a standard, and Google's Gemini can upload the same files, so a skill you write once can be added to other apps that support the format. Each app runs it in its own way. Where your playbook lives today:

Practice

Save the bakery rules as a skill, with one good example. In ChatGPT, paste everything below the second --- line into your project instructions.

---
name: cake-enquiry-replies
description: Use when drafting a reply to a customer asking about a cake order or collection.
---
# Cake enquiry replies
1. Answer only from the cake FAQ.
2. Ask for anything missing: the exact date and the cake details.
3. Never promise a cake. Staff confirm availability.
4. Style: friendly, short, no emojis.
5. End with a two-line summary for staff.

## Example
Customer: Can I collect a cake on Saturday at 3pm?
Reply: Hi! Collection is between 2pm and 6pm, so 3pm works. Custom cakes need at least 48 hours' notice. Which Saturday do you mean, and what cake would you like? Our team will check availability and confirm.

You've got it whenTwo different enquiries come back in the same shape and tone, and you didn't retype the rules.

Read

• What are skills? Claude Help Center · about 3 min Which plans have skills and how Claude decides to load one.
• Create & manage skills for Gemini Apps Google · about 8 min The Gemini steps, including uploading a SKILL.md.
• Claude Skills are awesome, maybe a bigger deal than MCP Simon Willison · about 8 min An independent writer on why a skill is "a Markdown file" and why that matters.

How it works

Without RAG, the AI answers from what it learned in training, which can be out of date or wrong for your business. With RAG, a search step runs first. AWS's explainer puts it simply: the model checks an authoritative source outside its training data before it answers. The model isn't retrained; it just gets the right page to read.

The search can match exact words, meaning, or both; Microsoft calls the mix hybrid search. You don't need to set any of that up. In Claude, a project holds your files (free accounts get up to five projects), and on paid plans Claude switches to RAG automatically when a project gets too big to read in one go. ChatGPT Projects hold files too: five per project on Free, 25 on Go and Plus, 40 on Pro. Google's Gemini Notebook, formerly NotebookLM, is built to answer from the sources you add.

The idea comes from a 2020 research paper by Patrick Lewis and colleagues at Facebook AI, now Meta.

Practice

Paste this into a chat, or add the FAQ to a project and ask the question there.

Answer the customer using only the FAQ below.
After your reply, list the FAQ line numbers you used.
If the FAQ doesn't cover something, say so. Don't guess.

FAQ (edition B):
1. Custom cakes need at least 48 hours' notice.
2. Collection is between 2pm and 6pm.
3. Staff confirm availability before an order is accepted.
4. We don't add discounts by message.

Customer: Can I collect a birthday cake on Friday at 4pm?

You've got it whenEvery claim in the reply matches a line you can find in the FAQ yourself.

Read

• What is retrieval-augmented generation? IBM Research · about 7 min The "open book" explanation and where the idea came from.
• RAG for projects Claude Help Center · about 3 min RAG in an app you can open today. Paid plans.
• RAG and generative AI Microsoft Learn · about 8 min The deeper version: word, meaning and hybrid search.

How it works

Memory is separate from the chat itself. Claude's memory is on by default on Free, Pro and Max, and it lists what it keeps under Topics, each one editable or deletable. ChatGPT shows a memory summary and saved memories you can correct or delete; OpenAI notes the controls vary by plan and region. In Gemini, anyone can add Instructions for Gemini (Settings & help, then Personal Intelligence); its memory of your details needs you to be over 18, on a personal Google Account with a Google AI plan, and is US only for now.

One detail trips people up: deleting a chat doesn't remove a memory made from it in Claude, and may not in ChatGPT. Both help pages say so. Delete the memory itself.

For a one-off chat that shouldn't add to memory, use a private chat: incognito in Claude (every plan), a temporary chat in ChatGPT, or Temporary Chat in Gemini. OpenAI says temporary chats don't create or update memories, though they can still use existing ones unless you choose Unpersonalized first.

Practice

Ask your AI what it knows, then open the memory settings and compare the two.

What do you remember about me? List each item on its own line.
Then tell me which items you'd use when drafting replies to customers.

You've got it whenYou've read the list, kept your reply style, and deleted anything about a customer or anything you'd rather it didn't know.

Read

• Memory in ChatGPT OpenAI Help Center · about 7 min The memory summary, deleting a memory, and why deleting a chat isn't enough.
• Use Claude's chat search and memory Claude Help Center · about 17 min Skim "View and manage your memory". Long; the team-plan parts can be skipped.
• Gemini adds Temporary Chats and new personalization features Google · about 4 min How Gemini uses past chats. From 2025; menu names have changed since.

How it works

Think of it as the AI's desk. IBM calls it the model's working memory. Everything for this reply sits on that desk, and what doesn't fit gets cut or summarised. That's why a long chat can lose your first instructions.

Bigger isn't automatically better. Anthropic's engineers describe "context rot": as the window fills, the model gets worse at recalling what's in it. Chroma tested 18 models and found answers grew less reliable as the input got longer, even on simple tasks. Anthropic's rule of thumb is the smallest set of useful information that gets the result.

Sizes vary by app and plan. On paid Claude plans, Claude's help page lists each model's window, from 200,000 to 1,000,000 tokens; the help page puts 200,000 tokens at about 500 pages. A big window still rewards a tidy desk.

Practice

Start a new chat with only three things: the job card from skill 1, the current FAQ (edition B) and the customer's message. Leave out last week's chats and the old FAQ.

[Paste the job card]

[Paste FAQ edition B]

Customer message: Can I collect a birthday cake on Friday at 4pm?

You've got it whenThe reply uses the current collection hours, 2pm to 6pm, and follows your job card.

Read

• What is a context window? IBM · about 10 min The plain definition, and what happens when you run past it.
• Effective context engineering for AI agents Anthropic · about 13 min Why context is "a critical but finite resource". The first third is the part for you.
• How large is the context window on paid Claude plans? Claude Help Center · about 2 min Real sizes per model. Numbers change often.

How it works

Anthropic's prompting guide says to be specific about the output format, and that examples are one of the most reliable ways to steer it. So name every field, say what goes in each, and say what to write when something is missing, such as "unknown".

Some apps help you move the result on. In Gemini, a reply that contains a table has an Export to Sheets button that saves it as a new Google Sheet. In Claude or ChatGPT, ask for the table and copy it across.

Developers can go further. Tools like Claude's structured outputs force the reply into a fixed format, called a schema, so other software can read it. You don't need that to start; it's what to ask for if you hand the job to a developer later.

Practice

Add this to the end of your prompt or playbook.

Return a table with these rows, in this order:
Requested day | Time | Date known? (yes/no) | FAQ lines used | Missing details | Draft reply | Needs staff? (yes/no)
Write "unknown" for anything the message doesn't say.

What a good answer looks like:

You've got it whenYou can paste the table into your sheet without editing, and "Date known?" says no.

Read

• Anthropic's prompting guide: control the format of responses Anthropic · about 3 min for that section Say the format, show an example, say what to do rather than what to avoid.
• Export responses from Gemini Apps Google · about 2 min One click from a table in Gemini to a Google Sheet.
• Structured outputs Claude docs · about 13 min What a forced format means, for the developer you hand off to.

How it works

Anthropic draws the line this way: a workflow follows paths set in advance, while an agent picks its own path, and workflows give you predictability for well-defined tasks. A repeat job like the bakery's is usually a workflow. The n8n docs add that every workflow needs at least one trigger to decide when it runs; they call each step a node.

Which one do you need?

Zapier's free plan gives you 100 tasks a month and two-step Zaps: one trigger and one action. So keep the first version small. The trigger is a new enquiry, and the action is an AI draft saved where staff will see it; review and sending stay with a person. An AI step in Zapier counts as one, three or five tasks depending on the model, and Zapier says the default models can be used for free.

Practice

Map one job on paper with four boxes. The bakery's:

1. Trigger: a new enquiry arrives by email.
2. AI step: draft a reply from the FAQ and fill in the table.
3. Review: a staff member checks the draft and the availability.
4. Action: the staff member sends it.

You've got it whenYou can point to the one box where a person says yes, and it comes before anything goes out.

Read

• Key concept glossary n8n · about 6 min Short, plain definitions: workflow, node, trigger.
• AI by Zapier: easily add AI steps to your workflows Zapier · about 9 min How an AI step sits inside a normal workflow, with examples.
• Workflow automation: definition, tutorial and tools Zapier · about 15 min The plain idea of a workflow before any AI is added.

How it works

Anthropic's guide to evaluations says to make the pass rule specific and measurable, and to use cases that look like your real work, edge cases included. "Sounds good" isn't a rule. "States the collection hours from the FAQ and promises nothing" is.

Developers run hundreds of cases and let software or another AI do the grading. You can start by hand: five cases, one pass rule, a tick or a cross for each. If you use an AI to grade, treat its score as one more answer to check.

Practice

Run each case in a fresh chat with your playbook. Every case passes only if the reply (a) matches FAQ edition B, (b) asks for missing details instead of guessing, (c) promises and sends nothing, and (d) keeps the style. Any promise is an automatic fail.

You've got it whenIt passes all five, and still passes after you change the playbook.

Read

• Define success criteria and build evaluations Anthropic · about 16 min Specific, measurable pass rules and cases that look like real work. Read the first half.
• Your AI Product Needs Evals Hamel Husain · about 17 min A practitioner's case for looking at real outputs and writing simple checks first. Some code.