CAGE sets the AI up before it does anything. Fill the fields, then copy a ready-to-paste block plus the all-important Align plan-check. Model-agnostic — paste into ChatGPT, Claude, Gemini, or a custom GPT. Nothing leaves your browser.
C
Context — all of it
load everything it would guess
Everything the AI would otherwise have to guess, loaded so it does not have to. The two layers travel together; add any domain facts and data too.
Completeness check:
Maturity: L0 loads none → L6 keeps engineered, reusable context blocks (a library others draw on). If you paste the same context twice, save it as a block.
A
Align — the plan-check
highest-leverage move
Before the AI produces anything, make it play back its understanding and its plan — like checking a capable new hire's plan before they spend two days building the wrong thing. Skip it and it charges off confidently in the wrong direction. 30 seconds here saves 30 minutes unwinding.
Maturity: L0 lets it run and is surprised → L6 treats the plan-back as a reliable gate and delegates with confidence. In an agent, this is "plan mode": approve the plan before you hand over control.
G
Goals — what "done" looks like
decision utility, not completeness
Not "write a report" but the decision this output must serve. Start general and let the AI help sharpen it — Goals are granular and bidirectional.
Tip: not sure of the criteria? Add to the Align step: "propose precise success criteria for this, then wait for my go."
E
Examples — show, don't just tell
a growing data layer
One or two samples of good — and ideally one of bad — calibrate quality faster than any description. Internal (your past work) or external (standards). Keep the good ones: every example you save becomes a future reference.
Your CAGE block
Paste this at the start of your chat (or as the system prompt / custom-GPT instructions). Run the Align step first — do not let it produce until you have approved the plan.