# How I run it

*Going further*

- Every setup in this guide, next to the bigger version I run.
- The real files my agents work from, cleaned.
- One link that hands every setup to your agent, and the order I'd build it all again in.

5 min read

This guide is a starter version of how I actually use AI. I run the same setups every day, across several agents and at a bigger scale. You don't need any of the bigger versions to get the value. They're here so you can see where each setup goes when you need more, and so you know the starter versions come from something that's in use.

## The short version

I send one request in plain words, and a main agent works out what's needed and does most of it. When I switch helpers on for a big job, one builder does the work and reports back, while planning, design and fact checks stay with the main agent. Every agent reads the same rules, the same memory and the same skills. Then it comes back to me, and nothing goes outside without my yes.

## Each setup, starter and full

| Setup | Starter version in this guide | How I run it |
|---|---|---|
| Ground rules | A short list in your standing instructions | One contract file every agent reads before it starts, with a short note for each app that points back to it |
| A memory that works | A one-page note, updated at the end of a session | A store of facts every agent reads through a read-only connection, where each fact has to quote its source |
| Fresh sessions | A handover note when a chat gets long | A search over thousands of past sessions, so any agent can find an earlier conversation instead of asking me again |
| Check the facts | Sources, a second pass and a fixed check | A rule that every fact is checked against a source before it reaches me, and scripts that fail a build when something's wrong |
| Straight answers | Standing instructions against flattery | The same lines in the contract, plus "challenge a shaky assumption with reasons" |
| Write like you | A voice file built from your writing | A voice guide every agent reads before writing in my name, grown from the changes I make to drafts |
| Save what works | A skill as a note | 33 shared skills in one folder, each with a dated changelog, which any agent fixes when it finds a better way |
| Routines | A morning brief or a weekly review | Routines that run while I'm asleep, each saved switched off until I've seen a run |
| Test your setup | Ten jobs and a marking prompt | Checks that run before anything ships, and a test I got wrong and published a correction for |
| More than one agent | One job per agent, and a hand-off note | Several bots with their own jobs, one builder at a time, and a label on every message between agents saying it isn't my approval |

The counts come from my own setup as of 2 November 2026, and the test results are on my results page (/results/).

## The actual files

These are real and cleaned. I've taken out names, paths, product names and anything to do with other people's work, and every other word is as my agents read it.

- My profile, an excerpt How my agents are told to talk to me and work with me. `real/profile-excerpt.md`
- The skills rule When an agent has to fix a skill or write a new one, and how it records why. `real/skills-rule.md`
- A routine's instructions A daily routine, with the fixes from its first run in. `real/link-sweep-routine.md`
- A hand-off note A real one, plus the packet my agents fill in for a big hand-over. `real/handoff.md`

## Hand it to your agent

Every setup in this guide is in one place on GitHub, at github.com/joelbrilliant/how-i-use-ai (https://github.com/joelbrilliant/how-i-use-ai). If your AI tool can read web pages, paste this into a new chat, and it'll walk you through the lot:

**Set me up from the guide**

```
Read https://github.com/joelbrilliant/how-i-use-ai, starting with SETUP.md. Walk me through setting myself up, one setup at a time. Explain each one in a line, show me exactly what you'll add, and wait for my yes before you change anything.
```

This guide tells you not to let an agent follow instructions it finds on the web, so the setup file is written the other way round. It's a list of setups for your agent to offer you, one at a time, and it asks your agent not to install, connect or send anything. Read SETUP.md (https://github.com/joelbrilliant/how-i-use-ai/blob/main/SETUP.md) yourself first if you'd like to see what it'll do.

If your tool can't read web pages, open the link and copy what you need, or use GitHub's Download ZIP button.

## What I'd do first

If I lost the lot tomorrow, this is the order I'd rebuild it in. It's my order, and yours might differ.

1. Ground rules, starting with draft only.
2. The memory note. Twenty minutes, and every answer after it gets better.
3. One skill, for the job I repeat the most.
4. A fixed check for the thing I can least afford to get wrong.
5. Ten test jobs, so I could tell whether any of the rest was helping.
6. One routine, once I trusted the brief.
7. Everything else, only when I could name what it was for.

The setups in this guide come from my own. It isn't a client case study.

If you want this set up for you or your team, that's what I do (/work-with-me/).
