Why Claude Gave You a Great Answer on Monday and a Flat One on Tuesday

Post 1 in a new series about my book, Claude AI Bible (7 Books in 1).

Have you ever asked Claude for the same thing twice and got back two answers that barely look related?

Monday it wrote something you were pleased with. You copied it, you used it, you thought: right, I've got this. Tuesday you typed what felt like the same request, and what came back was flat. Generic. Slightly off. Nothing you can point at and fix. Just not it.

If that has happened to you, you are not the only one.

A while ago I asked my readers one question: what is your number one frustration with AI? Over a hundred of you wrote back, and thank you for that. I expected "I don't know how to use it." That is not what came. The answers clustered, and they kept circling the same thing, in almost the same words: it gives me something different every time.

That answer is why this book has 242 pages and one spine.

The one sentence the whole book rests on

Here it is, from the first page of the Introduction:

Claude is not unreliable. Your inputs are.

That sounds harsh for about two seconds. Then it turns into the best news you will hear all week. You cannot rewrite the model. You can rewrite your request, and it takes about ninety seconds.

Think of it like this. Every new chat is a brilliant assistant on their first morning in the job. Quick, willing, reads faster than you do. And they remember nothing. Not your work, not your customers, not what "good" looks like to you.

On Monday, without noticing, you gave that assistant a lot. You mentioned who the text was for. You pasted an example. Somewhere in the conversation you said "keep it short and don't be salesy." On Tuesday you opened a fresh chat, typed one line, and expected the same standard.

Same assistant. Different brief. Different result.

Or, the way the book puts it: cooking from a recipe gives you a meal you can count on. Throwing ingredients in a pan gives you a surprise. Sometimes a good one.

The six-part checklist

The fix is a checklist you run every time, so nothing important gets left out by accident.

Mine has six parts. I call it CROFTC, and it stands for Context, Role, Objective, Format, Tone, Conditions.

• Context: the background. Who you are, what you are working on, why it matters.

• Role: who you want Claude to be while it works. Ask for feedback as a cautious investor and you get red flags. Ask as an enthusiastic mentor and you get next steps. Same question.

• Objective: the one main thing you want done. "Help me with my email" is a wish. "Write a reply declining this meeting politely" is an objective.

• Format: the shape you want back. A table with three columns. An email with a subject line. Under 200 words.

• Tone: how it should sound. "Friendly nudge: your payment is due Friday" and "Notice: payment required by Friday" carry the same fact. Only one sounds like you.

• Conditions: the rules and the limits. What to avoid, what to include, what to never do.

Six things. One page. You can learn it in an afternoon and use it for years, because it is not tied to any button in any menu.

Here is the difference in practice.

What most people type:

Write a newsletter about our spring sale.

The same request, briefed:

Context: I run a small garden centre. My newsletter goes to about 400 local customers, mostly aged 50 and over, who have bought from me before. Role: You are writing as me, the owner, not as a marketing department. Objective: Get people to come in during the first weekend of April. Format: One email, under 200 words, with a subject line. Tone: Friendly and plain. Like I am talking to a regular over the counter. Conditions: No hype words. No exclamation marks. Mention that the sale is only on bedding plants.

You are not writing a longer prompt to be clever. You are writing it so Monday's version and Tuesday's version are the same version.

Want the six parts as a one-page cheat sheet, with more than 500 prompts you can copy straight into Claude? I keep both in a free toolkit here: entrenexus.com/claude-ai-bonus-pack. No cost, and you can leave the list whenever you like.

What that turns into over 242 pages

CROFTC is the Foundations chapter. It is the first thing in the book and the one part I ask you not to skip, because everything after it is you pointing that single skill at bigger and bigger jobs. Seven books, easiest first.

Book 1, Chat and Projects. How to set Claude up once so it starts every conversation already knowing your work, instead of you re-explaining yourself daily.

Book 2, Prompt Engineering. Each of the six letters taken deeper, plus the Prompt Repair Method. When a result disappoints, you do not start over and hope. You ask which of the six elements caused it, change one thing, and compare. That loop is the difference between a frustrated beginner and someone in control.

Book 3, Code and Build. You describe a small tool in plain words and watch it appear. A tip splitter, a habit grid, a savings tracker. No code, and I mean that literally.

Book 4, Cowork. What changes when you stop sending messages and start handing over a whole job, with checkpoints where a human still decides.

Book 5, Skills and Agents. Here you meet the Tool Ladder: Prompt, Project, Skill, Agent. Four rungs, each doing more on its own than the one below. Once you can see the ladder, you stop guessing which one a job needs.

Book 6, Automation. Safe flows that run without you. Also a chapter called Avoid Irreversible Actions, which tells you what not to let run unattended. There is one optional advanced section in this book, and it is labelled as optional right there in the contents.

Book 7, Monetization. The order is value first, proof second, offer third. No income screenshots, no passive income while you sleep. The chapter is called Monetization Without Hype and it means it.

Three things run through all seven:

More than 500 ready prompts, numbered #001 to #505. The numbering runs straight through the whole book, so a prompt you liked three books ago is one number away. Treat them as a shelf you reach for, not a reading assignment.

Seven capstone projects. One at the end of every book. Twenty to thirty minutes each, built on your own real work, not a made-up example. Reading about a workout makes you informed. Doing it makes you stronger.

Dated boxes for anything that can change. Screens get redesigned. Features get renamed. Plan limits move. So every version-specific detail in the book sits in a labelled box stamped with the date it was true, kept apart from the method. One glance tells you how fresh that piece is. The method has no expiry date. The screenshots do, and I would rather say so than pretend.

What it will not do for you

It will not make you a developer. It will not hand you a business. The capstones take real time, and if you skip them you have read a book instead of gaining a skill.

And if what you want is a guided tour of every menu in the interface, there are other books that do that well. This one is built the other way round: method in the main text, buttons in the dated boxes. That was a deliberate bet, and the bet is that in two years the method will still be standing.

What is coming next in this series

I will take one piece at a time: setting Claude up so you never start from zero, the six letters one by one with fill-in templates, building a small tool without code, delegating a whole task, saving a method so you never explain it twice, connecting Claude to what you already use, and the honest version of turning it into paid work.

Before the next post

Take one thing you asked Claude for last week that came back not quite right. Write it again with all six parts. Five minutes, and if it does nothing for you, you have lost five minutes.

If it does work, the Foundations chapter is where the rest of it lives.

Claude AI Bible (7 Books in 1) is on Amazon here: amazon.com/dp/B0H9SXCQGX

The rest of this series.

Each of these takes one piece of what is above and goes further into it.

• Claude or ChatGPT? What Each One Is Actually Better At. The honest split by job, what each one is worse at, and why moving tools does not stop the drift.

• Stop Re-Explaining Yourself: Set Up a Claude Project Once. The workspace that holds your context between conversations, in five steps, with a worked example.

• Build a Small Tool in Plain English (No Code, No Kidding). Describe a small tool, watch it appear beside the chat, and test it before you trust it.

• The Tool Ladder: Prompt, Project, Skill, Agent. Four rungs, two questions that pick the right one, and why the top rung is not the prize.

Tomasz