Track 08 · How it is set up
A harness. Rules your AI reads first.
A harness is everything you put around an AI so it works the same way every time: the rules it reads, the methods it follows, what it remembers, the guards that stop it and the switches you hold. This page teaches track 08, using the setup Ru's own AI runs on every day.
Why your AI needs one
Clever is not the same as safe.
An AI on its own is a gifted stranger. It starts every chat blank, fills gaps its own way and, once it can send or delete, can do real harm quickly. A harness fixes each of those.
It forgets.
Every new chat starts empty. A harness is the first thing it reads, so you never explain yourself twice.
It guesses.
With no rules it does what seems likely. Written rules make the likely thing your thing.
It acts.
When an AI can press send, a polite rule is not enough. Something has to stop it, whatever it decided.
The five layers
Five layers. Each one plain.
From the softest to the firmest. The first three are words the AI reads; the last two work whether it listens or not.
- 01
Rules
A short page the AI reads at the start of every job: who you are, how you like things, and the few things it must never do.
In Ru's setup Drafts only: nothing is sent, posted or published except by a person. Nothing is thrown away; old versions are archived. A number without a date is not treated as a fact.
- 02
Skills
One short written method per job, kept in its own file. The AI reads only the names and one line on each, then opens the one that fits. Track 09 teaches you to write them.
In Ru's setup A skill for checking an outside tool before it is installed, one for how a page must look, one for building a website. Try the picker below.
- 03
Memory
Facts and lessons kept between chats, so a mistake made once is not made again. Each one carries the date it was learned.
In Ru's setup Short notes like "check the screen by looking at it, not by reading the code", written the day a check by code alone let a bad page through.
- 04
Automatic guards
Small scripts that run on their own before the AI acts, and block anything that breaks a rule. Track 10 calls them hooks.
In Ru's setup A guard that refuses to delete files, so the rule "nothing is thrown away" holds even on a bad day.
- 05
Switches
Every feature the AI adds gets an on and off switch in one place you control. Turning something off never needs the AI's help.
In Ru's setup One list of features with a switch beside each, shown in his own app, read by the AI before it uses any of them.
Try it · layer 02
It reads the names. It opens one.
Your AI does not read every skill it has: that would fill its memory before the work began. It scans a list of names, like the one below, and opens only the one that matches the job. Type a job and watch.
Or try , , .
- 0
- names read
- 0
- skill opened
Opened
Nothing yet.
- Loading the list
A plain word match, written for learning. A real AI matches meaning, not just words.
Next
Write your own. Then check theirs.
Track 08 ends with your own rules page and one skill of your own. Tracks 09 and 10 show how to check a skill or plugin someone else wrote, before you let it in.