Start here · First agent
Your first agent¶
This walkthrough gives an AI a folder of its own, one small job, and clear rules — the whole loop, start to finish, with nothing at risk.
1. Make the folder¶
Create a new folder anywhere easy to find. Inside it, make two more:
ai-practice/
├── input/ ← copies of a few non-private files
└── output/ ← where the AI puts its drafts
Put 2–3 copies of harmless files in input/ — some notes, an article you saved, anything you'd like summarized or reorganized.
2. Open your agent in that folder¶
In a folder agent (like Claude Code), open or start it inside ai-practice/. If it asks for permissions, that's normal — but it should only need this folder.
3. Give it the job¶
Work only inside this folder. Read the files in input/ and write
a clear one-page summary of them to output/summary.md. Before
writing anything, tell me your plan in two sentences and wait
for my OK. Don't change anything in input/.
4. Watch and approve¶
It should tell you its plan, wait, then write one file. If it tries to do more, say stop. You're the supervisor.
5. Check the work¶
Open output/summary.md and ask yourself:
- Did it only use what's in
input/? - Did it invent anything?
- Did it touch anything it shouldn't have?
- Is the summary actually useful?
That's it — you supervised an agent¶
Everything bigger on this site is this same loop with more steps: clear job, limited access, plan first, human check.
Next: install a playbook — a job your agent saves and remembers, so you never have to explain it twice.