AI Guides › Workbench
By Nigel Guy · 6 min read
You ask Claude whether your plan is sound, and it tells you the plan is promising, then lists a few "things to consider". It feels like feedback. It is mostly agreement with a polite garnish, and you walk away more confident and no better informed.
This is a known tendency, not a personal failing of yours. Anthropic's research paper "Towards Understanding Sycophancy in Language Models" found that five AI assistants showed sycophantic behaviour across several tasks. It also found that human preference data tends to reward responses that match the user's views, so models can learn to agree. You will not fix that with a magic word. You fix it by asking for a different job.
The rule: never ask Claude whether you are right; assign it a specific adversarial job, with a defined output, before it is allowed to answer.
Short capitalised labels such as EXPOSED, BLIND SPOT, STEELMAN and AUTOPSY circulate online as "codes" for Claude. They are not commands. We could not find any Anthropic documentation that defines them, and Claude has no hidden switch behind them. They work, to the extent they work, because the word carries an instruction Claude already understands. A label alone is ambiguous, so the blocks below spell out the job each label stands for. We also could not verify the "twenty more codes" claim, so we have not reproduced a list. Anthropic's own guidance is to be clear and direct and to give context and success criteria, which is exactly what the blocks do.
| Block | Job it assigns | Free? | Best for | Catch |
|---|---|---|---|---|
| Exposed | Reflect back the assumptions and pattern in how you framed the problem | Works on any plan; nothing to buy | Early framing, when you suspect you are asking the wrong thing | Claude only sees what you typed, so it infers; make it label inferences |
| Blind spot | Name the better question | Same | Decisions where you have fixated on one option | Can invent drama; ask for one question, not ten |
| Steelman | Build the best case against your position, then give a verdict | Same | Opinions, pitches, strategy | The verdict can slide back to flattery; demand a stated condition for changing your mind |
| Autopsy | Assume the plan failed and trace why | Same | Launches, migrations, hires, budgets | Produces plausible stories, not probabilities |
No prices apply: these are prompt patterns and work in the free tier of Claude as well as paid plans. Check the £ price of any paid plan on Claude's pricing page at checkout if you want longer or heavier conversations.
What it forces: Claude reads your own wording back as evidence of assumptions you have not stated.
You are a sceptical analyst. I will describe a situation and I want you to examine how I have framed it before you answer anything.
Context: [DESCRIBE THE SITUATION IN YOUR OWN WORDS, MESSY IS FINE]
Task:
1. List up to four assumptions my wording implies but never states.
2. For each, quote the phrase of mine that suggests it.
3. Mark each as "stated by me" or "inferred by you", and give your confidence in the inference (low, medium or high).
4. Do not advise yet.
Before you reply, check that every inference is tied to a quoted phrase. If I have given too little to infer anything, ask me up to three questions instead of guessing.
Fill in: the situation, in your own unpolished words. Polished briefs hide the framing you want exposed.
What it forces: a better question, not a better answer to the one you asked.
You are a senior advisor who has seen many people make this kind of decision.
My current question: [YOUR QUESTION]
Background: [RELEVANT FACTS, CONSTRAINTS, DEADLINE]
First, give the single question I should be asking instead, in one sentence, and say in two sentences why mine is the weaker one. Then answer your better question briefly, and finish by answering my original question in the light of it.
Rules: do not offer more than one replacement question. If you need a fact from me to choose it, ask first. Flag anything you are assuming. Check before answering that your replacement question genuinely differs from mine and is not a rephrasing.
Fill in: your actual question and the real constraints.
What it forces: the strongest opposing case, and a verdict that has to survive it.
You are a debate coach who must argue the side I disagree with, at full strength.
My position: [YOUR POSITION IN ONE OR TWO SENTENCES]
Why I hold it: [YOUR REASONS]
1. Write the strongest honest case against my position, in the form its smartest supporter would give it. No strawmen, no hedging.
2. Identify which of my reasons is weakest.
3. Give a verdict: still right, partly right, or wrong. Name the specific evidence that would change that verdict.
Constraints: do not soften step 1 to be polite. Do not flip your verdict merely because I push back later; change it only if I supply new facts. If you cannot argue the other side without knowing more, list what you need.
Fill in: the position and your reasons. The last constraint matters; without it, a second message of "but I think..." often undoes the work.
What it forces: a failure story told backwards, ending at a fixable weak point.
You are a post-incident reviewer. Treat the plan below as having already failed, six months from now.
Plan: [YOUR PLAN, WITH DATES, BUDGET AND PEOPLE]
What success looked like: [YOUR DEFINITION]
1. Write the three most plausible causes of failure, most likely first.
2. For the top cause, walk backwards from the failure to the earliest decision that made it likely.
3. Name the one weak point I can strengthen this week, and one cheap test that would show whether it is real.
Constraints: use only facts I gave you; label guesses as guesses. Do not give probabilities you cannot justify. Ask for missing details before writing. Self-check: each cause must point to a concrete part of my plan.
Fill in: a plan with real detail. A vague plan gets a generic failure story.
You can chain them, but the output grows long and the later blocks get thinner. A workable order is Exposed, Blind spot, then your real question, Steelman, Autopsy. Paste each block's job under a numbered list and add: "Complete them in order, keep each section under 150 words, and mark anything you inferred about me." For anything important, run them as four separate messages instead; you can respond to each before the next.