AI Guides › Workbench
By Nigel Guy · 7 min read
You ask ChatGPT about your own plan and get back something tidy, encouraging and pitched at a general reader. It feels useful because it agrees with you. It has skipped what you keep circling without saying, the answer an expert would want and the ways the plan fails. "Code words" such as EXPOSED, IQ 200 and AUTOPSY get passed around as fixes. They do help, but not for the reason people give.
The rule: the capitalised word is a label, not a switch. ChatGPT has no hidden command called AUTOPSY. What changes the answer is the instruction you write after the label, so write that part properly.
| Label | What it actually asks for | Cost at time of writing | Best for | The catch |
|---|---|---|---|---|
| EXPOSED (pattern read) | Name the recurring theme, worry or assumption in how you've described something | Free; works on any ChatGPT plan, including Free | Long chats where you've been going round in circles | It can only see this chat, plus whatever Memory surfaces. Its read on you is an inference, not a fact |
| IQ 200 (full-depth answer) | Drop the beginner framing; give the expert-level answer with trade-offs | Free; any plan | Questions where you already know the basics | It does not make the model cleverer. It removes simplification, and sometimes adds jargon and false confidence |
| AUTOPSY (premortem) | Assume the plan has already failed, then explain the most plausible causes | Free; any plan | Testing a plan before you commit time or money | It will always find causes, because you told it the plan failed. That is a story, not a forecast |
None of these needs a paid plan.
A label marks a change of mode and gives you a handle to reuse later ("AUTOPSY again, harder"). The plain-English instruction underneath does the rest. The premortem is a real technique: Gary Klein described it in Harvard Business Review in 2007 as a team imagining a project has already failed and generating reasons why. Asking ChatGPT to do it borrows the method, not any evidence that the method predicts your outcome.
What it forces: before answering, the model has to report back how you've been framing the problem: the assumption you repeat, the option you keep avoiding, the word you keep using.
How to use it: run it after several messages, not as your opener, because one message has no pattern to find. Check Settings → Personalization → Memory. If "Reference chat history" is on, ChatGPT may draw on past conversations too, which can sharpen the read or drag in something stale. A temporary chat set to Unpersonalized limits the read to what you've typed in that chat.
Fill in [SUBJECT] with the thing you've been discussing.
EXPOSED. Pause before giving any advice. You are acting as a candid
thinking partner who reads framing, not just content.
Look back over everything I have written in this conversation about
[SUBJECT]. Report:
1. The assumption I seem to treat as settled.
2. Any option or risk I have mentioned once and then dropped.
3. Words or phrases I keep repeating, and what they might signal.
For each point, quote or closely paraphrase the message it comes from.
Label each one either "from your words" or "my inference".
If there is too little in this chat to see a pattern, say so plainly
instead of inventing one. Keep it under 200 words.
Before answering, check that every point traces back to something I
actually wrote.
What it forces: no beginner padding. It cuts the definitions, names the trade-offs and pitches the answer at a specialist.
How to use it: the label is theatre. What does the work is telling ChatGPT what you already know. If you don't know the field, skip it: an expert-sounding answer you can't check is worse than a plain one you can.
Fill in [QUESTION], [WHAT YOU ALREADY KNOW] and [DECISION THIS FEEDS].
IQ 200. Treat me as a practitioner, not a newcomer.
Question: [QUESTION]
What I already understand: [WHAT YOU ALREADY KNOW]
The decision this answer will feed: [DECISION THIS FEEDS]
Give the answer you would give a specialist colleague:
- Lead with your direct answer in two sentences.
- Then the main trade-offs, including the ones that cut against my
apparent preference.
- Then the conditions under which your answer would flip.
Skip definitions of anything listed under what I already understand.
Mark any claim you are unsure of as "uncertain" rather than smoothing
it over. If the question is missing something you need, ask me one
question before answering.
Before you reply, reread your answer and remove any sentence that only
restates the question or offers general encouragement.
What it forces: the model argues against your plan instead of polishing it, working backward from a failure you've told it has already happened.
How to use it: give it enough detail to fail realistically: timeline, money, people and the step you're least sure of. Ask for an early warning sign for each cause, so you end up with something to watch for.
Fill in [PLAN], [TIME HORIZON] and [WHAT FAILURE MEANS].
AUTOPSY. Act as a sceptical reviewer running a premortem.
My plan: [PLAN]
Imagine it is [TIME HORIZON] from now and the plan has failed, where
failure means: [WHAT FAILURE MEANS].
Working backward from that failure:
1. List the five most plausible causes, most likely first.
2. For each, give the earliest warning sign I could spot, and roughly
when it would appear.
3. For each, give one cheap check I could run this week.
Stick to causes tied to details I gave you. Do not pad with generic
risks such as "poor communication" unless the plan specifically points
to them. Separate what follows from my description from what you are
assuming about my situation.
This is a stress test, not a prediction: do not state or imply odds of
failure. If key details are missing, list them and ask before you
finish.
Check before replying: is every cause specific to this plan?
Pattern first, then depth, then premortem. The pattern read can show you've been asking the wrong question, the depth answer tackles the right one, and the premortem tests the plan that comes out of it. To run them as one message, fill in the four bracketed fields and send it after you've described your situation:
Run three passes, in this order, with a heading for each.
Pass 1 – EXPOSED: from my earlier messages about [SUBJECT], name the
framing I keep returning to, citing where it shows up.
Pass 2 – IQ 200: answer [QUESTION] at specialist depth. I already know
[WHAT YOU ALREADY KNOW]. Direct answer first, then trade-offs.
Pass 3 – AUTOPSY: treat [PLAN] as having failed. Give the most plausible
causes, most likely first, each with an early warning sign.
Tag anything about me that you inferred rather than read. If a pass
can't be done well with what you have, say so instead of filling space.
You told the model the plan failed, so it will always find causes, however sound the plan is. A long, confident list doesn't mean the plan is weak, and a short one doesn't mean it's safe. Treat the output as a list of things to rule out, then rule them out with real information: numbers, a customer conversation, a small test. A cause that keeps coming up when you re-run the prompt in fresh chats deserves a closer look. EXPOSED has a milder version of the same problem: "you seem anxious about money" is a reading of your wording, not a diagnosis.
Going round in circles: EXPOSED. Answers feel basic, in a field you can check: IQ 200. About to commit money or time: AUTOPSY. Big decision in a long chat: all three, in order.