AI Guides › Workbench
By Nigel Guy · 7 min read
Most people ask an AI model the same way: paste the work, type "thoughts?", and take whatever comes back. The default reply is polite, balanced and padded, so it feels helpful while changing nothing. Prompts that "feel illegal" are not secret tricks. Each one changes the job the model is doing: from agreeing to attacking, from summarising to teaching, from advising to comparing, from describing to producing a procedure, and from rewriting everything to fixing one thing.
The rule: never ask for "feedback" or "help". Name the job, name what a good result looks like, and name what the model must not do.
Anthropic's research on sycophancy defines it as model responses that match a user's beliefs over truthful ones, and found that human preference data can reward agreement. In plain terms: if you paste a plan and ask whether it is good, the model is under mild pressure to say yes. Anthropic's own prompting guide offers a useful test for any prompt: show it to a colleague with minimal context on the task, and if they would be confused, the model will be too. Each prompt below is built to pass that test. It states a role, a goal, an output shape and a stop condition.
| Prompt | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| 1. Hostile review | Attacks a plan, then repairs it | Free tiers work; paid plans only matter for long documents | Plans, pitches, launch ideas | Can invent flaws to look tough |
| 2. Smart beginner | Explains a topic without jargon or condescension | Free tiers work | Unfamiliar fields, contracts, reports | Simplifying can quietly drop a caveat |
| 3. Practice gap | Compares your habits with how experienced people in the field tend to work | Free tiers work | Career, craft, routines | Generic unless you supply your own details |
| 4. Repeatable system | Turns something you did once into a procedure | Free tiers work | Anything you do weekly | Bakes in your mistakes if the original was weak |
| 5. Weakest-part patch | Finds one weak point and rewrites only that | Free tiers work | Drafts you mostly like | Needs your own judgement on what "weakest" means |
On cost: all five run on a free plan of Claude or ChatGPT. At time of writing Claude's Pro plan is listed at $20 a month (about £15 at time of writing; check the £ price at checkout, as tax and exchange rate move it). Pay only if you hit usage limits or work with very long documents. Plans and limits change often, so check the vendor's pricing page.
The source version of this idea was a one-line roast. A one-liner gets you theatre: dramatic insults, no priorities. Give the attack a structure and a repair.
You are a sceptical reviewer whose job is to find the reasons this plan could fail. You are tough but accurate: you never invent problems.
Context: [WHO_I_AM_AND_WHAT_I_AM_TRYING_TO_ACHIEVE]
Plan: [PASTE_PLAN]
Constraints I cannot change: [BUDGET_TIME_OR_OTHER_LIMITS]
Do this in order:
1. List the five most likely ways this fails, ranked by likelihood times damage. For each, quote the part of my plan it comes from.
2. For each, say what early warning sign I would see.
3. Give one concrete fix per failure that fits my constraints.
4. Finish with the single change that would improve the plan most.
If you need information to judge any point, ask me before answering rather than guessing. Before replying, check that every criticism points at something actually written in my plan.
Fill in: your context, the plan and the fixed constraints. The quote requirement is what stops invented flaws.
"Explain it simply" produces baby talk. The useful instruction is to say what you already have (intelligence) and what you lack (vocabulary).
Act as a patient teacher. I am intelligent and a fast learner but have no background in [TOPIC]. Do not talk down to me and do not use jargon without defining it the first time.
Material: [PASTE_TEXT_OR_NAME_THE_CONCEPT]
What I need it for: [DECISION_OR_TASK]
Give me:
1. A three-sentence summary.
2. The five terms I must know, each with a plain definition and one everyday comparison.
3. How the pieces connect, in a short paragraph.
4. Three things people commonly get wrong.
5. Two questions I should now ask an expert.
Mark anything you are unsure of as uncertain. If the material is missing context you need, ask first. Before replying, check that no undefined term remains.
"What would the top one percent do?" has no data behind it. No model has a ranked list of the top one percent of your field. Reword it as a comparison against widely documented practice, and make it use your details.
You are an experienced practitioner in [FIELD] giving a candid review of my habits.
My current routine, tools and output: [DESCRIBE_WHAT_YOU_ACTUALLY_DO_WEEKLY]
My goal over the next [TIMEFRAME]: [GOAL]
Compare my approach with commonly recommended practice in this field. Return a table with columns: Area, What I do, What experienced practitioners typically do, Effort to change (low/medium/high), Expected benefit.
Then pick the three changes with the best benefit for the effort and say which to start with this week.
Do not claim statistics or name studies unless you are certain they exist; say "commonly advised" instead. If my description is too thin to compare, ask me questions first.
You are a process designer. Convert what I did below into a procedure someone else could follow without asking me anything.
What I did and the result I got: [DESCRIBE_OR_PASTE_THE_PROCESS_OR_OUTPUT]
How often I will repeat it: [FREQUENCY]
Tools available: [TOOLS]
Output:
1. A one-line purpose.
2. Inputs needed before starting.
3. Numbered steps, each beginning with a verb, with a note where a judgement call is needed.
4. A short checklist to confirm the output is right.
5. Three likely failure points and what to do about each.
Do not add steps I did not describe unless you label them "suggested addition". Ask about any gap rather than filling it with a guess.
You are a line editor. Find the single weakest section of the text below and rewrite only that section.
Text: [PASTE_TEXT]
Purpose and reader: [WHAT_IT_IS_FOR_AND_WHO_READS_IT]
"Weak" means: [UNCLEAR_OR_LONG_OR_UNSUPPORTED_OR_OFF_TONE]
Steps:
1. Name the weakest section and explain the problem in two sentences.
2. Provide the rewrite, matching my tone and length within about 10 per cent.
3. Leave everything else untouched, and say so.
Before replying, check that your rewrite keeps every fact I gave and adds none.
Defining "weak" matters: without it the model chooses its own standard.
Chain them: 1 finds the flaw, 5 patches it, 4 captures the result.