AI Guides › Workbench
By Nigel Guy · 7 min read
You paste in a plan, an email or a draft, and Claude tells you it is a strong start with a few small suggestions. It feels like feedback and it is mostly politeness. The default reply to a loose request is a pleasant, balanced, middle-of-the-road answer, and you can go weeks without noticing that you never asked for anything else.
The rule: Claude answers the question you actually asked, so ask for the job you need done (critique, plain language, a verdict, a gap check), and name it in one short phrase at the start.
A note on the title. Social posts call these "cheat codes". They are not secret codes and there is no hidden mode. Each phrase is an ordinary instruction that tells Claude which of several valid answers you want. Anthropic's own prompting guidance says the same thing at length: Claude responds well to clear, explicit instructions, and if you want a particular kind of behaviour you should ask for it rather than expect it to be inferred. These five are our selection, not an official list.
A vague request has many acceptable answers, and Claude picks a safe, agreeable one. A phrase that states the job removes most of that choice. Anthropic's guide describes Claude as a brilliant new colleague who lacks your context, and suggests a test: show your prompt to a colleague with no background and see whether they would be confused. "Thoughts?" fails that test. "Tell me the three weakest points in this" passes it.
Anthropic also publicly treats sycophancy, meaning telling people what they want to hear rather than what would help them, as a problem it measures and works to reduce in its models. That helps, but it is not a guarantee, and how blunt a reply feels varies between models. Your phrase still does the steering.
| Phrase | What it changes | Best for | Catch |
|---|---|---|---|
| "Be blunt: what is wrong with this?" | Praise first becomes problems first | Drafts, plans, pitches | Can overcorrect into nitpicking |
| "Argue the other side." | One-sided agreement becomes the strongest opposing case | Decisions you have already half made | Opposing case is not the verdict |
| "Explain it plainly, as if I have no background." | Jargon becomes plain steps | Contracts, reports, new subjects | Plain can mean oversimplified |
| "Verdict first, then reasons." | A survey of options becomes one answer | Choosing between two options | A confident verdict can hide thin evidence |
| "What am I not asking?" | Answers your question becomes questions you missed | Planning, research, early ideas | Some gaps will be irrelevant |
Cost: all five work on any Claude plan, including the free one, which has usage limits. Claude Pro is listed at $20 a month ($17 a month billed annually) on Anthropic's pricing page, and the £ price depends on your region and VAT, so check the £ price at checkout. Phrases cost nothing extra.
Use it when you want the truth rather than applause. Give Claude the material, the standard it should meet, and permission to be unpleasant.
You are a sceptical reviewer with no stake in my feelings. Review the [DOCUMENT_TYPE] below against this goal: [GOAL_AND_AUDIENCE].
Be blunt. Start with the three most serious problems, ranked, each with a quote from the text and a specific fix. Then list anything minor in one line each. Only after that, name one thing that works and why.
Do not soften the problems, and do not invent problems to look thorough. If something is fine, say it is fine. If you need information I have not given (for example who will read this), ask me before reviewing.
Before you answer, check that every criticism points at something actually in the text.
[PASTE_TEXT_HERE]
Fill in: the document type, the goal and audience, and the text.
When you have already decided, ask for the best case against, not a balanced list. This is the steelman move: the strongest version of the view you disagree with.
I am leaning towards [YOUR_DECISION] because [YOUR_REASONS]. Act as a thoughtful advisor who thinks I am wrong.
Give the strongest honest case against my decision in no more than five points. For each, say what evidence would prove the point right or wrong, and what I could check this week. Then say in two sentences which of your points worries you most.
Do not caricature my position and do not argue for the sake of it. If you lack facts about my situation, list the questions you would need answered first.
Self-check: would someone who holds the opposing view agree this is a fair statement of it?
Fill in: the decision and your reasons.
This is for clarity. The trick is to say who you are and what you will do with the answer, so Claude can choose the right level.
Explain [TOPIC_OR_PASTED_TEXT] to me in plain British English. My background in this area is [NONE / SOME: DESCRIBE]. I need to understand it well enough to [WHAT_YOU_WILL_DO_WITH_IT].
Format: a two-sentence summary, then the key points as a short numbered list, then any term I must know with a one-line definition. Use one everyday comparison at most, and only if it is accurate.
Do not simplify by leaving out something that would change my decision. Flag anything you are unsure of rather than smoothing it over.
Fill in: the topic or text, your background, and what you need it for.
When you are stuck between options, a survey of pros and cons delays the decision. Ask for a ruling.
I am choosing between [OPTION_A] and [OPTION_B] for [CONTEXT]. My priorities, in order, are [PRIORITIES]. My constraints are [BUDGET_TIME_OTHER].
Give your verdict in the first line: A or B. Then give the three reasons that carried the decision, one thing that would make you change your mind, and the biggest risk of your pick.
Do not answer "it depends" without saying what it depends on and which way you would lean. If a fact I have not given would flip the answer, ask me for it first.
Fill in: both options, the context, your ranked priorities and constraints.
The risk in any request is the thing you did not know to mention. This phrase turns Claude from answering to auditing.
I am planning [PROJECT_OR_QUESTION] and so far I have thought about [WHAT_YOU_HAVE_COVERED]. My aim is [OUTCOME].
List the questions I should be asking but am not, grouped as: things that could stop it, things I am assuming without evidence, and things other people involved would raise. Rank the top five by how much damage ignoring them could do. For each, say how I would find out the answer.
Keep to what is relevant to my aim, and say plainly if you are guessing about my situation.
Fill in: the project, what you have covered, and the outcome.
Start with what you are about to do. Sending something out: phrase 1. About to commit: phrase 2, then 4. Confused: phrase 3. Early planning: phrase 5. You can stack two in one message, but not five; a prompt that asks for blunt critique, a verdict and a gap check at once produces a muddle.