Framing your AI: the complete guide
Your AI is brilliant for five minutes and a disaster on the sixth. It forgets what you said yesterday, reinvents what already worked, states figures with total confidence. You spend more time supervising it than moving forward, and you start to think the tool simply is not ready.
Framing your AI means installing three pieces in your project: an instruction sheet it re-reads at every opening, written rules it has to follow, and automatic barriers that stop it before a blunder. One hour to install, no code to write, and it finally becomes predictable.
Why it is unpredictable without a frame
The official documentation says it plainly: every conversation begins with an empty memory. Your project, your prohibitions, last week’s incident — none of it survived yesterday’s closing.
Until something is written down, you are the one playing the part of the memory. You repeat, you supervise, you catch things. A frame moves that load off your shoulders and onto files, which never get tired of it.
Piece one — the sheet it re-reads every time
It is a simple text file placed in your project folder, which the AI reads automatically at every opening. Inside, three blocks: what the project does, how you want to work, what is forbidden.
Two traps to avoid. First: writing too much — the documentation recommends staying under 200 lines, because a long sheet is followed less closely than a short one. Second: writing wishes (“be rigorous”) instead of checkable orders (“show me the plan before changing more than one file”). The difference between the two has its own article: how to give instructions to a coding AI.
The full mechanism is covered in why your AI forgets your instructions.
Piece two — the written rules
The rules are the part of the sheet that changes behaviour. The five that pay the most, in order:
The plan before the action — it shows you what it intends before doing it. One task at a time — it flags other problems instead of “fixing” them in passing. Do not touch what works. Delete nothing without asking. Never “done” without proof.
Each rule is an order checkable in one second. If you cannot verify that a rule was followed, it is badly written.
Piece three — the automatic barriers
A written rule can be ignored — rarely, but it happens, and the day it happens on a deletion, you get nothing back.
The barriers are of another nature. They are checks that fire before the risky action and prevent it, without asking the AI’s opinion. Wiping files in bulk, writing a secret key into the code, opening the passwords file: stopped before they exist.
A rule asks for cooperation. A barrier does not ask.
The install, in order
One hour, four stages. Place the sheet and check it: close everything, reopen, ask “what does this project do?” — right answer, piece validated. Add the rules and test one: ask for a change touching several files, the plan must come first. Install the barriers and run the risk-free test: ask it to delete everything, it must get blocked. Finish with the outer net: a version tool and a copy elsewhere, because no frame replaces a backup.
What the frame does not do
It does not make the AI smarter and it does not write your strategy for you. It makes its work predictable: same instructions, same limits, same checks, every session. That is exactly what separates the people who save time with these tools from those who lose it — the measured gap runs from zero to 81% productivity, on the same tool.
> pick the project back up, carry on with the pricing page