Three Rules Anthropic Engineers Actually Use When Prompting Claude
- Chris Gore
- 3 minutes ago
- 5 min read
Three rules Anthropic engineers actually use when prompting Claude, stop thinking in prompts, start building skills, and why the output compounds over 30 days
Chris Gore | Updated 2026
I sat in on a talk by Anthropic's own engineers and walked away feeling like I had been using Claude wrong for months. The fix is embarrassingly simple. No technical background required. Done in ten minutes. These are three rules that most people who use Claude every day have never heard of, because the people who built Claude have not made a big song and a dance about them. Until now.
Why Most People Get Inconsistent Outputs From Claude
Here is what most people do. They open Claude, type a question, get an answer and close the chat. The next time they need the same thing they open Claude again, type a slightly different version of the same question and get a slightly different answer. Then they wonder why the outputs feel inconsistent.
Think about it like this. Imagine hiring the best contractor in the world. And every single morning you hand them a completely blank brief. No context. No history. Nothing about what they built yesterday. Just a job. Off you go. You are leaving an enormous amount on the table because the real value of that contractor is not just what they can do. It is what they learn about the work you are asking them to do. You have the smartest assistant in the world and every day they come back to a blank slate again and again. That is not a Claude problem. It is a prompting problem.
Rule One: Stop Prompting Claude. Start Prompting Skills.
This is a mental shift more than a technical one. Anthropic engineers do not write new prompts for everything. They build what they call skills. Folders that package up repeatable tasks so that Claude knows exactly how to handle them every single time.
Most people, myself included for a long time, think in prompts. One task, one message, one response. But most of what you do in business is repetitive. You are drafting similar emails, reviewing similar documents, creating similar content. And you are writing the same brief over and over again each time as a prompt. Instead of writing that prompt every single time, you write it once properly and save it as a skill. Next time you come back you just use the skill. Same quality. Every single time.
The mental model that made this click for me: Anthropic built the phone. Your job is to build the apps that sit within the phone. Most people are typing in the browser every single time. They are trying to build the app from scratch every single time. Stop thinking in prompts. Start thinking in skills. That is where everything else builds from.
Rule Two: A Skill Has Three Layers. Most People Stop at Two.
Layer one: the description
This is what Claude checks every time to decide whether the skill applies to what you are asking. If it is vague, Claude guesses. If it is specific, Claude knows. That is the difference between an assistant that anticipates and one that shrugs its shoulders and makes something up. Get this specific.
Layer two: the instructions
The step-by-step process. The playbook. This is where most people stop. And honestly, if you stop here you will get decent results. But decent is not why you are building skills. You are here for exceptional.
Layer three: the tools
Scripts, reference files, external connections. Attach your brand guidelines, your pricing schedule, your standard contract terms. Attach all of that into the skill and it changes the quality of the output completely. Now Claude is not inferring. Claude knows because you put it in there. This is where a skill stops being a smart prompt and becomes a scalable system. Anthropic engineers go all in on layer three. Most people have never touched it.
You do not need to be a developer to get value from layer three. A reference document is enough. Your brand voice document. Your pricing guide. Your contract terms. Attach them and watch what happens to the output quality.
Rule Three: Your Skills Should Get Smarter Every Time You Use Them.
Most people run a skill, get an output and move on. The skill stays the same. The output stays the same. No improvement over time.
Anthropic engineers do something different. After every output they ask one question before moving on: is this a one-time fix or should this skill live forever? The answer is always live forever. They update the skill. Add the new rule. Add the edge case. Add the example.
Here is why this matters. Every session with Claude starts from scratch. There is no memory. No continuity. But if you update the skill every single time you use it, the skill carries that memory. Not Claude, the skill. Day one your skill produces a decent piece of work. By day thirty, after thirty updates, it produces something that feels uncannily tailored to exactly what you want. That is not magic. It is compounding. Thirty small improvements consistently applied.
Most people write a prompt once and never touch it again. The Anthropic engineers treat it like a living document. Every session is a chance to make the next session better. That is the whole game. For more on how to deploy Claude properly in a business context, read the guide to how to set up Claude CoWork for your small business.
Want the Weekly Thinking on AI and Business Productivity?
Work Isn't Working — the newsletter by Chris Gore. One email a week on what is actually working. No fluff. Free.
Join the newsletter > spor-group.net/work-isnt-working-newsletter |
Frequently Asked Questions
What is a Claude skill?
A Claude skill is a saved, reusable task that packages up a repeatable workflow so Claude handles it consistently every time. Unlike a one-off prompt, a skill has a description, instructions and tools — reference files and external connections that give Claude the context it needs without you rewriting the brief each time.
Why do my Claude outputs feel inconsistent?
Because every Claude session starts from a blank slate. Without skills, you are writing a slightly different prompt each time and getting a slightly different output. Skills solve this by giving Claude consistent context and instructions every single time.
What are the three layers of a Claude skill?
Layer one is the description — what Claude checks to decide if the skill applies. Layer two is the instructions — the step-by-step playbook. Layer three is the tools — reference files, scripts and external connections. Most people stop at layer two. Anthropic engineers go all in on layer three.
How do you make a Claude skill get better over time?
Update it after every use. Add the edge case. Add the new example. Add the rule you discovered during this session. Claude has no memory between sessions but the skill does. By day thirty of consistent updates the output is uncannily tailored to exactly what you need.
Do I need technical skills to build Claude skills?
No. The most valuable layer three addition — reference files — requires nothing technical. Attach your brand guidelines, pricing schedule or contract terms as a document and Claude will use them as context for every output the skill produces.
Comments