One for how an agent thinks, one for making something new, one for fixing it so it stays fixed. Keep this card next to you while you work.
Whose experience it thinks through: the expert it acts like, its filters, its always and never.
Ask yourselfWhose eyes should this be seen through?
What you are really after and why: the goal, what done looks like, when it should ask you.
Ask yourselfCould I write the goal, done when and ask if in three lines?
What is true right now and where it lives. One home per fact, read fresh, never an old copy.
Ask yourselfIs it reading the live source, or something it saved last week?
What it can do and inside what limits: its tools, its skills, what needs your OK first.
Ask yourselfCan it actually act on this, and is it allowed to?
Knowledge sits under all four. It is the model, the same floor everyone has. Your edge is the four letters on top.
What it is, who it is for and why. Get grilled on it, then see it as a diagram before anything is built.
Ask yourselfCan I see it on one page, and has it survived the hard questions?
Connect the data, tools and accounts it needs. Prove each one with a live read.
Ask yourselfWhat does it need to reach, and have I proven it can?
Load the LIFE layers. Teach it how you want it done: your lens, your intent, the facts, the limits.
Ask yourselfDoes it know how I would do this, not just what to do?
Is the problem big enough? Weigh the value against the time, money and energy it costs. Look for lopsided upside.
Ask yourselfIs the payoff many times bigger than what it takes to build and run?
Ship it, test it against real use, and improve it in rounds.
Ask yourselfIs it live, tested on real work, and better than last week?
Zoom out first. Is this the real problem or a symptom? Is it worth fixing? Get the whole story, not a screenshot.
Ask yourselfIs this the broken traffic light, or just the car in front of me?
Write what was supposed to happen: the goal, done when, and the one thing that should come back to you.
Ask yourselfWould two people read my goal the same way?
Replay it to the exact moment it went wrong. Treat the agent's excuse as a clue. Name the letter that broke.
Ask yourselfWhich LIFE letter broke?
Fix it in the system, not the prompt. Borrow before you build. Remove the need, add a wall or a checker; a new instruction comes last.
Ask yourselfCan I take away its ability to get this wrong?
Prove it on the real case, have it reviewed, guard the whole class of problem, make it run without you, save the lesson where it loads.
Ask yourselfIf I stepped away for a month, would this stay fixed?
Make something new. Its Intel step is where the LIFE layers get loaded.
The context any agent needs to do the job your way.
When something goes wrong, LIFE is the checklist for finding where.