This is a three-prompt chain. The first two write an emergency preparedness guide. The third one tears it apart.

That third prompt found the hole sitting in the middle of the finished guide. It never tells you how to actually produce any heat. It rules out burning anything indoors, which is correct, and then it just stops — leaving you a plan whose real answer is “eventually you leave,” without ever saying that out loud.

Then it explained why the hole was there:

It was easy to leave out because the honest answer is unsatisfying and costs money, and retention advice is free and sounds prudent.

The solution is actually given right in that feedback. Give the model auditing as a separate job, and it will tell you things it will never volunteer while it is still busy trying to be helpful.

Why “Be Critical” Does Nothing

You have probably tried tacking be critical or point out any weaknesses onto the end of a prompt. I have. It produces a polite paragraph at the bottom that reads like a disclaimer and changes nothing about the work above it.

The reason is simple enough. While a model is producing something it is invested in that thing being good, so criticism written in the same breath as the work is just the author defending the work. You want the audit happening in a turn where the model has no stake in what came before and its only job is finding problems.

So this runs as three prompts instead of one, and they work better the further apart you keep them.

Pass 1: Make It Think Before It Writes

The first prompt does not write the guide. It produces a failure analysis — what breaks in what order, this household’s specific weak points, the one decision that actually matters, and what they can stop worrying about entirely.

Skip this and every recommendation in your guide gets generated alongside the situation rather than derived from it. That is a decent part of why AI prep guides all read the same regardless of who they were supposedly written for.

You are producing a one-page failure analysis for a specific household facing a specific
emergency. This is not a guide and not advice. It is the reasoning layer a guide will
later be built on.

THE SITUATION
Threat: multi-day winter grid failure, 3 to 7 days, sustained below freezing, roads
intermittently impassable.
Household: two adults, two children aged 4 and 8, one 40lb dog.
Home: second-floor apartment, no balcony, all-electric heat and cooking, one interior
stairwell, municipal water.
Already have: one vehicle at half a tank, ordinary kitchen supplies, no generator, no
yard, no stored fuel.

PRODUCE FOUR THINGS, nothing else:

1. FAILURE ORDER
   What stops working, in the order it stops, from minute one to day seven. Be specific
   to all-electric and second-floor. Note which failures cascade into others.

2. THIS HOUSEHOLD'S FOUR WEAK POINTS
   Vulnerabilities created by who lives here and where they live, that a generic
   checklist would not catch. One sentence of consequence each.

3. THE DECISION THAT MATTERS MOST
   Name the single highest-stakes decision this household faces, state when it must be
   made, and give the observable signals that force it either way.

4. WHAT IS ALREADY FINE
   What this household does not need to worry about, and why. Be generous here.

Under 500 words total. Plain sentences. No preamble, no summary, no encouragement.
Do not offer solutions - only the analysis.

It came back with things a generic checklist does not produce. The dog got flagged as forcing more cold exposure than anything else in the plan, on the grounds that it needs to go outside several times a day and cannot use a bucket. The 4-year-old got flagged too:

Small bodies lose core temperature faster than adults and cannot reliably report it, so the adults will still feel fine while the child is already in trouble.

Pass 2: Now Write It

The second prompt takes that analysis and builds the guide on top of it. The instruction doing the heavy lifting is that every recommendation has to trace back to something the analysis raised. No analysis, no bullet point.

Below is a failure analysis for a specific household. Write their preparedness guide
from it.

Every recommendation must trace to something in the analysis. If the analysis did not
raise it, leave it out. Where you reference a weak point or the key decision, do it in
plain language rather than by number.

[PASTE PASS 1 OUTPUT HERE]

WRITE FOUR SECTIONS:

1. THE GO BAG - about 350 words
   Three tiers, labelled: gets us out alive, gets us back to normal, keeps us going.
   Every item in exactly one tier, each with a one-clause reason drawn from the analysis.
   Flag anything this household needs that a generic list would miss.

2. MAKING THIS HOME HOLD - about 350 words
   Ordered by what to do first. Address the failure order directly: what to do before the
   heat is gone, and what changes once it is. Cover water, warmth, cooking, and the
   stairwell.

3. STAYING REACHABLE - about 300 words
   Four tiers by what infrastructure has to survive: (a) needs no power at all, (b) runs
   on stored battery, (c) needs regional infrastructure intact, (d) needs only satellites.
   Describe the capability and the buying criteria. Categories only - no brands, no model
   numbers.

4. THE FIRST 48 HOURS - about 400 words
   Chronological, grouped by hour blocks, starting the minute the power drops. Every line
   is an action performed, written as an imperative. Put the key decision from the analysis
   at the hour it has to be made. No explanation - just the moves in order.

RULES
Calm and direct. Short sentences. Assume intelligence, assume zero experience. Second person.
Invent nothing you cannot stand behind: no dosages, no frequencies, no legal claims about
what is permitted where, no prices, no model numbers.
Do not pad. Finish early rather than fill.
Plain prose and simple lists. No tables.

This is the long one. Give it room and let it finish before you touch anything.

Pass 3: The Red Team

Now we really refine it further. Paste the finished guide into a fresh turn and hand it four specific things to hunt. Specific is the whole game — vague instructions to criticize get you vague criticism.

Below is a finished preparedness guide. Your job is to break it.

Read it as someone who will have to follow it at 3am, in the dark, cold, with two
frightened children and no second chance.

[PASTE PASS 2 OUTPUT HERE]

FIND FOUR THINGS:

1. ASSUMED CAPABILITIES
   Every instruction that quietly assumes a tool, a skill, a working phone, a store being
   open, or a spare pair of hands this household was never confirmed to have. Quote the
   instruction, name the assumption.

2. STEPS THAT FAIL IN THE DARK
   Anything that only works with light, warmth, dexterity, or calm. Say what it degrades
   into when those are gone.

3. DEAD WEIGHT
   Items or steps earning less than they cost in space, time, or attention. Say what to
   drop and what to carry instead.

4. THE MISSING PAGE
   The one thing this guide does not cover that its reader will most likely need. Name it
   and say why it was easy to leave out.

Then rewrite ONLY the weakest single section of the guide, corrected. Label it clearly as
the replacement.

Be blunt. Do not compliment the guide. Do not restate it. If something holds up, say
nothing about it.

Those last two lines do more work than they look like they do. Do not compliment the guide kills the throat-clearing paragraph where the model reassures you about your own work. If something holds up, say nothing about it stops it padding the response out to feel thorough.

What It Actually Caught

It caught the guide contradicting itself. The guide had already established that power gets cut deliberately across the whole area, then cheerfully told the reader to go top up their fuel tank.

It caught invented precision, which is something that is always challenging to safeguard with AI:

“Final walk of the perimeter, thirty seconds.” Thirty seconds is a number the guide invented to sound decisive.

It killed two of its own recommendations. Dripping the faucet got cut for spending stored water the same guide treats as scarce, and a whistle got cut for solving a problem this household does not have.

And it produced the best line in the whole thing, on how to check a child for cold once your own hands have stopped being reliable instruments:

Check that child against the back of your own neck rather than your hands. Cold hands lie.

This Runs On Anything You Sell

The preparedness guide is just what this one is aimed at. The red team pass has nothing to do with prep.

Aim it at a recipe book and it finds the step assuming two free hands. Aim it at a business checklist and it finds the task assuming a tool the reader never bought. Aim it at a course outline and it finds the module quietly depending on knowledge from a module you placed later.

If you sell how-to material, this is a QA pass you can run across your whole catalog. Here it is with the brackets open.

Below is a finished [WHAT YOU MADE - guide / checklist / course outline / recipe book].
Your job is to break it.

Read it as someone who will have to use it [AT THE WORST REALISTIC MOMENT - at 3am in the
dark / mid-shift with no help / on a phone in a parking lot], with [WHO OR WHAT THEY ARE
RESPONSIBLE FOR] and no second chance.

[PASTE YOUR CONTENT HERE]

FIND FOUR THINGS:

1. ASSUMED CAPABILITIES
   Every instruction quietly assuming a tool, a skill, a working system, an open business,
   or a spare pair of hands the reader was never confirmed to have. Quote the instruction,
   name the assumption.

2. STEPS THAT FAIL IN [THE DEGRADED CONDITION THAT APPLIES TO YOU - the dark / a rush /
   without supervision]
   Anything that only works with light, time, dexterity, or calm. Say what it degrades into
   when those are gone.

3. DEAD WEIGHT
   Items or steps earning less than they cost in space, time, or attention. Say what to drop
   and what to keep instead.

4. THE MISSING PAGE
   The one thing this does not cover that its reader will most likely need. Name it and say
   why it was easy to leave out.

Then rewrite ONLY the weakest single section, corrected. Label it clearly as the
replacement.

Be blunt. Do not compliment the work. Do not restate it. If something holds up, say nothing
about it.

The line to customize is the worst-realistic-moment one. That single sentence does most of the lifting, because it hands the model a concrete lens instead of asking it to be critical in the abstract.

Where It Gets Annoying

Three separate turns, and Pass 2 runs long, so this is not a thirty-second job. Budget for it.

Pass 3 will occasionally hand you a complaint that is technically correct and practically useless. You are still the filter. It generates candidates, you make the calls.

And it is genuinely uncomfortable to read. The first time it tells you the thing you just built has a hole in the middle, the instinct is to argue with it. Sit with that anyway, because that reaction is more or less the point.

One thing I will not oversell: this does not make the model smarter than it already was. It just stops it grading its own homework, and that turns out to be most of the problem.

Go Run It

Three prompts, in order, same chat. Fill Pass 1 with a household and a threat you actually care about, let Pass 2 write, then set Pass 3 loose on what it made.

Then go find something you already sell and run the red team pass on that instead. That is where it earns its keep. (Fair warning — it will find something. It always finds something.)