Most people still use AI like a faster search engine. Five stages take you from a single question to an agent that does a real piece of work every week without being asked.
- 1AskGet an answer
- 2ThinkFind the gaps in your idea
- 3TeachShow it how you like things done
- 4DelegateHand over a whole job
- 5AutomateLet it repeat on a schedule
You do not need to code for any of this. The first three stages happen in an ordinary chat with Claude or ChatGPT. The last two use Claude Cowork (or Claude Code or OpenAI’s Codex), which can work with files on your computer.
Stage 1: Ask
This is where everyone starts: you type a question, the AI answers. The message you type is called a prompt.
The fastest way to get better answers is to say what kind of help you want. Three details make the biggest difference:
- Role: who the AI should act as, for example a patient teacher
- Output: what it should produce, for example a short table
- Audience: who it is for, for example someone new to the topic
Try asking about the same subject for two different audiences and compare the answers:
Explain how a small business sets its prices, as if I were 10 years old.
Explain how a small business sets its prices, as if I had an MBA.
If an answer feels generic, the fix is usually to say who it is for.
Stage 2: Think
Here the AI stops being an answer machine and becomes a sparring partner that pushes back on your ideas. It is the best way to find blind spots before they cost you.
Here is my biggest challenge at work right now: [describe it]. Find my blind spots and my riskiest assumptions.
Pick one point from the reply and test it. Then ask the question most people skip:
Which parts of this work could I drop completely without hurting my goal?
Removing a task beats automating it.
Stage 3: Teach
Think of the AI as a capable new hire on day one. It is not bad; it just does not know how you work yet. Give it what you would give a new person:
- Context: the facts it needs
- Examples: what good and bad work look like
- Clarifying questions: let it ask before it starts
- Constraints: limits such as length or what not to touch
- Feedback: what to keep and what to change
- Checks: look at the result before you use it
This template covers most tasks:
Your task: [the goal]. Context: [the background and facts it needs]. Example: [a sample of good work]. Constraints: [length, tone, what not to touch]. Before you start, ask me 5 clarifying questions.
Use it on something you write often, like a post or an email. Answer its questions, give feedback until you like the result, then, in the same chat, save what it learned:
Turn what we did in this chat into a skill, so I never have to explain it again.
A skill is a saved set of instructions the AI can reuse. Next time, a short request like "use my writing skill to draft a post about [topic]" is enough. When it gets something wrong, correct it and ask it to update the skill.
A ready-made skill for writing LinkedIn posts, built on the method above. It includes a file called voice-example.md: paste a sample post into it before you upload, or just paste one in the chat when the skill asks.
Download the LinkedIn writing skill (ZIP)To add a skill to Claude, click Customize in the sidebar, then Skills, then the + button, and upload the ZIP as it is (do not unzip it). In Claude Code, unzip it into the .claude/skills folder in your home folder (~/.claude/skills).
Stage 4: Delegate
Now give the AI a meaningful unit of work, not a single answer. A good first job: the weekly business review. Every Monday you need to know what happened last week, but the answers are scattered across a sales spreadsheet, customer feedback and team updates.
Put those files in one folder, open Claude Cowork, add the folder and use the Stage 3 template:
Your task: prepare our weekly business review and an agenda for the leadership meeting. Context: this folder has sales data, customer feedback, team updates and last week’s review. Compare performance, find important changes and check last week’s action items. Example: follow the structure of last week’s review. Constraints: no more than two pages. Name the source file for every key finding. Flag anything missing. Save new files and never overwrite the originals. Before you start, ask me 5 clarifying questions.
Then check its work properly: compare the numbers with the spreadsheet, make sure every explanation is backed by the files, and confirm the original files are unchanged. When you are happy, save the process as a skill with the same prompt as in Stage 3.
A ready-made skill that recalculates every metric from the files, checks team updates against the data, labels guesses and missing information, and never touches the input files.
Download the weekly business review skill (ZIP)Stage 5: Automate
The last step turns the tested skill into an agent that runs without you. Two things make that possible:
- Tools: connect the apps the report needs (in Claude these are called connectors), so the agent reads live data instead of files you export by hand. Start with read-only access.
- A schedule: ask it to run the skill on a fixed day and time.
Create a scheduled task that runs my weekly-business-review skill every Monday at 8:00 AM.
Check the schedule, the time zone and the instructions once it is set, and run it once manually before you rely on it.
Use it on your own work
Pick one task that takes you 30 to 60 minutes every week and follow the same order: ask, think, teach, delegate, automate. Ask the AI to interview you about the task first:
Interview me about how I do [task], from start to finish. Ask one question at a time. Then help me write the goal, context, examples and constraints for a reusable skill.
Your expertise is knowing the job and checking the work. Your clarity is how well you explain the goal, context, examples and constraints. Add tools and schedules only once you trust the result, and keep checking in even when it runs on its own.
Quick answers
Do I need to code?
No. Everything here uses plain-language instructions and, for stages 4 and 5, a folder on your computer.
Is Claude Cowork free?
No. It needs a paid Claude plan.
What is the difference between a skill and an agent?
A skill holds reusable instructions for one job. An agent uses instructions and tools to carry out the job, and decides how along the way.
Does saving a skill mean the AI learned?
No. It saved a playbook it can read next time. The model itself does not change, so test the updated skill in a new chat.
How do I keep company data safe?
Use an approved tool, a dedicated folder with only what the task needs, read-only access and a separate place for drafts. Permissions protect data; telling the AI "do not change anything" is only an instruction.
How do I measure the return?
Time the task by hand, then time the AI version, including preparing the files, reviewing and fixing the result. The difference, multiplied across tasks and weeks, is the time you buy back.