Your old prompts probably still work. But that doesn't mean they are the best way to work with new models. When a model is more autonomous, excessive, repeated or contradictory instructions can become a brake: they consume more tokens, make the process slower and sometimes reduce flexibility.
The practical difference for a company is simple: a good prompt no longer tries to control every step. Define the expected outcome, provide the right context, set the hard boundaries, and leave room for the AI to resolve within that framework.
The 12 keys of the manual
Fewer rules, more clear goals
Instead of dictating each microinstruction, I described the desired outcome and expected work style. Hard rules are still necessary for security, formatting, or scope.
Take out the generic controls
Phrases like “check everything carefully” often trigger vague verifications. I asked for specific controls: compare prices, validate a table or review a specific field.
Define the scope
When the task is small, say so. “Change only this section” prevents the model from improving nearby areas that you did not want to touch.
I said the length explicitly
If you need a short answer, ask for it with a limit. Effort level controls how much you think, not necessarily how much you write.
Monitor progress notifications
Define if you want updates, when and in how much detail. In automations, too much intermediate text gets in the way.
Subagents only when they help
Delegating is useful for large and parallel tasks. In small tasks it can multiply cost and time without improving the result.
Give context by parts
Don't put everything in one giant prompt. The general thing goes to the system; the specific is loaded when the task needs it.
Eliminate repetitions
If a rule is already in a tool description, do not repeat it in the general prompt. Less duplication means less conflict.
Use real references
It is better to deliver the file, mockup, table or concrete example than to ask for something “modern” or “similar” in a vague way.
Try low effort
For well-defined tasks, the low or medium level may be enough. The best configuration is the cheapest one that maintains quality.
I reduced the reasoning, don't turn it off
For agents and tools, it is often safer to keep reasoning on at low effort than to turn it off completely.
The goal is the path
The modern prompt describes the entire goal, sets limits and provides references. It does not drag old controls that no longer contribute.
Practical examples from then and now
Don't write long comments. Do not create additional files. Check everything carefully. Use a second agent to validate.
Orient yourself in the existing code. Take its nomenclature, structure and density of comments. Compare the prices with the attached table.
Improve this.
Work only on the task described. Make minor routine decisions on your own. Do not modify adjacent areas.
The golden rule
A good prompt for new models answers five questions. If those five questions are clear, the AI needs fewer control instructions and can work better.
- What must exist in the end? The specific result you are looking for.
- What information or references should you use? Data, examples, real files or business rules.
- What hard lines should you not cross? Security, fixed format, privacy, reach or human decisions.
- What length and what format do you need? Brief, detailed, table, checklist, email, report or code.
- What requires control or delegation? What is reviewed, what is compared, and what can be divided into parallel tasks.
Describe the entire goal, set the limits, give quality references and let the model choose the most sensible path within that framework.
What does this mean for a company?
For a business in any market, writing better prompts is not a technical game. It is a way to reduce cost, avoid rework and obtain more useful answers for sales, administration, customer service, reports or internal documents.
The point is not that every employee memorizes theory. The point is to turn real processes into clear instructions: what to respond, what data to ask for, when to escalate a person, what not to promise, and how to measure whether the result is useful.
That's why at Ezelero we don't see prompting as a nice phrase written in a chat. We see it as part of the work architecture: rules, context, tools, limits and measurement.
Frequently asked questions
Do I have to rewrite all my old prompts?
No. But it is advisable to review long, repetitive or full of generic controls prompts. Those are the ones that tend to consume extra tokens the most.
Are shorter prompts always better?
Not necessarily. A good prompt can be short or long. The important thing is that you have fewer unnecessary rules and more task-specific context.
What mistakes cost the most money?
Lazy checks, unnecessary subagents, unlimited response lengths, repeated instructions, and giant system prompts that load context that the task doesn't need.
Is this just for Claude?
No. Although the manual focuses on Claude Opus 5, many rules apply to modern AI models that work with tools, reasoning, and more autonomy.
Do you want prompts and flows adapted to your company?
At Ezelero we convert real processes into assistants, automations and work rules with AI. First we review your operation; Then we define what should be automated.
The original PDF is available as a downloadable guide for your team.