Image generated with AI. Art direction: Grafica.
Several of the rules our AI agents follow go back to a dated mistake, written down so it would not happen the same way twice. Our repository holds 25 such lessons. These are three of our rules for working with AI, and the mistake behind each. Our post How we use AI at Grafica covers what AI may not do at all. These three are about how its work is checked.
1. Nothing is missing until someone has looked where it would be
On 7 August 2026 an inventory of our own folders was compiled. One section reported that a project had no copy on our local drives. The project was there. It lived on an external drive that the machine running the inventory could not see. The report had a warning for exactly that case, and the warning worked in one section and not in another.
The rule: before anything is reported as missing, the report says which machine looked, and whether that machine could see the place where it would be. Until then, «not found» describes the search, not the world.
2. A limit you remember is a limit you made up
On 19 August 2026, in one session, five written claims about GitHub’s programming interface turned out to be false once someone asked the interface itself. One of them, that a setting could not be changed through it, had turned a job the interface could do into a manual step that was not needed.
The rule: a claim that a tool cannot do something is checked against the tool before anyone acts on it.
3. A command nobody has run is a sentence about a command
On 15 August 2026 an instruction was handed over with a stand-in in it, a label in angle brackets where a real number belonged. It was pasted and run with the label still in it. Two days later another block carrying a stand-in failed as a whole, so none of its five lines ran, including the first, a routine upload of finished work.
The rule: an instruction handed to anyone, a person or another agent, carries real values only. Nothing in it is a stand-in: it runs as written.
What the three have in common
In each case a report sounded certain and described something other than the machine in front of it. Each one was written down as a lesson, and the rule now stands where the next agent reads it before it starts.
Three questions to ask of any report, from an agent or a person
- When it says something is missing: where did it look?
- When it says a tool cannot do something: did someone ask the tool?
- When it hands you a command: is there anything in brackets, and has anyone run it?
We work this way on our own systems. If you want a vendor whose reports say how they know, talk to us.
Gráfica built this survey as a WordPress plugin, and builds plugins like it for clients. Tell us what your site needs
