AI Guides › Workbench
By Nigel Guy · 6 min read
Someone posts a list of twenty-five short capital-letter "codes" for ChatGPT, you paste a handful into every chat, and the answers feel sharper for a day. Then they get stranger: shorter in places that need length, contrarian about things you did not ask about, and oddly stiff. The codes were never switches. ChatGPT has no hidden command layer that recognises "BLUF" or "AUTOPSY"; each one is an ordinary instruction wearing a nickname.
The rule: a code is only a label for one specific instruction you could have written out in a sentence, so use the few that fix a problem you actually have, and write each one out in full at least once.
The idea circulating online is a list of about twenty-five short phrases, each meant to override one default behaviour: padding, shallow answers, agreement with you, unchecked reasoning. I could not find any OpenAI documentation describing such codes, so treat the idea as community shorthand, not a product feature. What is documented is more useful. OpenAI's help centre says custom instructions let you share anything you want ChatGPT to consider, applied to all chats. OpenAI's own prompting guide says GPT models benefit from explicit instructions about how to do a task, and that reasoning models respond better to a clear goal than to a long list of micro-steps.
The mechanism behind the better-known failure, the yes-man, is also documented. In April 2025 OpenAI rolled back a GPT-4o update it said was overly flattering or agreeable, and said it had leaned too hard on short-term feedback such as thumbs-up ratings. It pointed users to custom instructions as a way to shape behaviour. So an "honesty" code is a reasonable thing to want. It is a request, not an override, and the model can still drift.
| Code (my wording) | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| BLUF | Answer first, reasoning after | None; custom instructions are on all plans at time of writing | Briefings, decisions, anything you skim | Can clip nuance if you also demand brevity |
| KILL CRITIC | Argues against your leaning | None | A real decision with a stake | Will invent objections if your situation is vague |
| FLAG GUESSES | Marks inference vs what you said | None | Advice about your own circumstances | Labels are self-reported, not audited |
| AUTOPSY | Assumes the plan failed, finds the weak point | None | Plans, launches, proposals | Gives a plausible story, not a probability |
Custom instructions have a length cap: the help centre gives 1,500 characters on Free and Go, 5,000 on Plus, Pro, Business, Enterprise and Education, as at time of writing. That cap is itself an argument against stacking.
Edits to custom instructions apply immediately, including to existing conversations, per the help centre.
Use it when you get three paragraphs of warm-up before the point.
You are a concise briefing writer. For the question below, put the direct answer in your first sentence. Then give the reasoning in no more than [NUMBER_OF_BULLETS] bullets, most important first. If the answer depends on something I have not told you, say so in the first sentence instead of guessing.
Question: [YOUR_QUESTION]
Context: [RELEVANT_BACKGROUND]
Before replying, check that the first sentence could be read alone and still be useful.
Fill in: the question, background and bullet count.
Use it on a decision you are already leaning towards. The aim is a fair opposing case, not a personality change.
You are a sceptical advisor who is paid to find the flaw in my current plan, not to please me. Here is the decision: [DECISION]. Here is what I am leaning towards and why: [MY_LEANING_AND_REASONS]. Here are the facts and constraints: [FACTS_AND_CONSTRAINTS].
Do this in order:
1. State my reasoning back in two sentences so I can correct it.
2. Give the three best arguments against my leaning, strongest first.
3. Name where I may be rationalising, citing my own wording.
4. Say what evidence would change your view in either direction.
Rules: do not invent facts about my situation. If a detail you need is missing, ask me before answering. If my leaning is actually sound, say so and say why. Self-check: would a fair-minded expert accept each objection as real?
Fill in: decision, your leaning, facts. Vague input gives vague objections.
You are an analyst who separates what I told you from what you assume. Topic: [TOPIC]. What I have told you: [MY_STATED_FACTS].
Answer the question [QUESTION]. Mark every statement about my situation as STATED (I said it) or ASSUMED (you filled a gap). Finish with a list of the assumptions that, if wrong, would change your advice most, and ask me to confirm or correct them. Do not present an ASSUMED point as settled.
Fill in: topic, your facts, your question.
You are a reviewer running a pre-mortem. My plan: [PLAN]. Timeline: [TIMELINE]. Resources and limits: [RESOURCES].
Imagine it is [DATE_IN_FUTURE] and this plan has clearly failed. Work backwards: list the most likely causes in order of likelihood, say which single cause you would fix first and how, and name one early warning sign for each. Base this only on what I have given you; if the plan is too thin to judge, ask me up to [NUMBER] questions first. Self-check: is every cause tied to something in my plan, not a generic risk?
Fill in: plan, timeline, resources, date.
Match the code to the failure, not to the list.
Stacking all twenty-five is worse than using none, for ordinary reasons. Instructions can conflict (be brief, but show your full reasoning; be sceptical, but be warm). A permanent contrarian setting makes ChatGPT push back on questions that did not need it. Your instruction budget is capped. And a long block of capitalised codes is harder for you to audit than three plain sentences. I have not tested stacks of any particular size, so I will not claim a number; the practical test is to add one code at a time and keep it only if the answers to the same question improve.