AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people who feel let down by AI have not met a weak model. They have met a default: a blank chat, no background, no files, no instruction to disagree, and a quick answer to a question that needed a slow one. It feels fine while it is failing, because the reply is fluent and confident and agrees with you.
The rule: before you judge an answer, check the five settings that decide its quality, namely memory, pushback, thinking, inputs and honesty about guesses.
This guide is written around Claude, because those settings are documented and I could verify them. The ideas carry over to other assistants, but the menu names will not.
Run it as a table. Each row is a sign you are still on the beginner setting, followed by the switch that replaces it.
| # | Sign | What it costs you | The switch |
|---|---|---|---|
| 1 | Every chat starts from zero | You re-explain your job, your goals and your style each time | Memory, a Project, standing instructions |
| 2 | It agrees with everything | A plan with a hole in it, praised | Ask for the case against, before the recommendation |
| 3 | You never choose how hard it thinks | Shallow answers to decisions with real trade-offs | Check the thinking setting for the model you are on |
| 4 | You ask the question without the material | A generic answer that fits anyone | Attach the files and numbers, and state what you are optimising for |
| 5 | You never ask what it is guessing | Invented specifics sit beside real ones | Ask it to mark every assumption |
Signs 1, 2 and 3 are about how the tool is set up. Signs 4 and 5 are the two nearly everyone is guilty of, because they are about habit, not menus: you give too little, and you do not ask it to be honest about what it filled in.
At time of writing, Anthropic's help centre says memory is on by default for Free, Pro and Max plans. On Team and Enterprise, owners control availability and it stays off until each member enables it. To check yours, go to Settings > Memory and look at the "Generate memory from chats" toggle. The same page lists what Claude has stored under "Topics", where you can edit or delete individual entries.
Three things to know:
For work that repeats, a Project is the better home. Projects are self-contained workspaces with their own chats, uploaded files and project instructions. Free accounts can have up to five. Paid plans add retrieval to handle larger amounts of material. Put your stable context there once: who you are, what you are working on, what good looks like.
A model that is asked "is this a good plan?" tends to find the good in it. The fix is structural. Ask for both sides argued properly before any recommendation, and say plainly that you want disagreement. You will see this built into the prompt below.
In Claude, extended thinking is switched from the model picker: click the model name next to the send button, hover over "Effort", and use the "Thinking" (or "Extended") toggle. Anthropic's help page says it raises token use and so you reach usage limits sooner, which is the trade-off for deeper reasoning.
The catch is that this varies by model. The same page says thinking cannot be turned off on Claude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1 or Claude Opus 5. On those, the control is about effort, not on or off. Model lists change often, so open the picker and read what you actually see. Spend thinking on decisions, analysis and anything with several constraints. Skip it for rewording an email.
One structured message handles both habits. Fill in the brackets, attach your files, and paste it.
You are acting as [ROLE: e.g. a sceptical finance advisor for a small UK business].
Background
- Who I am and what I am working on: [YOUR_SITUATION]
- What I am optimising for: [GOAL_AND_CONSTRAINTS]
- Attached material: [FILES_OR_NUMBERS]
The decision I am stuck on: [DECISION]
A good result is a clear recommendation I could defend to someone who disagrees with it.
Work in this order:
1. Read the attachments. If anything essential is missing or unclear, ask me up to five specific questions and wait for my answers. Do not guess missing facts.
2. Make the strongest honest case for each option, one after the other.
3. Only then give a recommendation, with the main reason and the main risk.
Rules
- Do not flatter me or soften bad news. If my reasoning has a gap, say so first.
- Mark every statement that rests on an assumption as [ASSUMPTION], and every figure not found in my material as [NOT FROM YOUR FILES].
- Do not invent numbers, sources or quotes.
Before you answer, check: did you use my attachments, did you argue both sides fairly, and did you mark every assumption? Fix anything that fails.
Fill in the role, your situation, the goal, the attachments and the decision. If you are on a model where thinking is optional, switch it on for this message.
Say you run a small design studio and cannot decide whether to hire a part-time freelancer or buy a premium software subscription to cut turnaround time. The beginner message is "should I hire or buy software?" The checked version sets the role, attaches last quarter's project hours, states that you are optimising for turnaround without raising fixed costs, and uses the prompt above. The model asks which tasks eat the hours, argues for each option, and flags that its estimate of freelancer rates is an assumption you must verify. You still decide. You now have something to decide against.
At time of writing, Claude's Free plan is $0, Pro is listed at $20 a month billed monthly or $17 a month on annual billing (about £15 and £13 at time of writing), and Max starts from $100 a month (about £75 at time of writing). Memory is included on all three, and so are projects, with Free limited to five. Check the £ price at checkout, since currency and tax change the figure.