RICO: a prompt you can decide with

If you ask an AI "recommend a good family car," you get a list of five SUVs with the usual pros and cons, no ranking and no recommendation. That is information. To decide, you still have all the work ahead of you.

Write the same task with structure and the type of answer changes: you get something you can decide on or present. I call that structure RICO. It has four components, always in the same order, and it works the same in ChatGPT, Claude, Gemini or Copilot. The letters come from the Spanish version of the method; in English they read Role, Instruction, Context, Output.

The four components

Letter	Component	Question it answers	What it includes	Typical mistake
R	Role	Who is the AI at this moment?	Experience, point of view and authority. "Senior financial analyst, 15 years explaining closes to management."	A decorative role ("you are an expert"). With no years, no specialty and no point of view, the role changes nothing.
I	Instruction	What does it have to do?	A clear verb, a narrow scope and the depth you expect.	A soft verb ("could you help me with...?") or two tasks mixed into one.
C	Context	What would a colleague need to know to do this well?	What the answer is for, who reads it, which data to use and which constraints to respect.	Leaving it out. It is the block people skip most and the one that changes the result most.
O	Output	What does "done" look like?	Format, structure, length and tone.	Skipping it. Without an Output, the AI picks the format for you, and it almost always picks long.

The same task, written two ways

How most people write it:

Recommend a good family car.

And with RICO:

R: You are a car advisor with 15 years of experience in Costa Rica. Direct, not a salesperson.
I: Recommend 2 specific cars, with your reasoning and the trade-offs.
C: Family of 4. I drive in the city every day and head to the mountains on weekends. Budget USD 35K. I prefer a hybrid. My last car was a RAV4.
O: A table with: model, why, what I give up, what it solves better than the other option.

The first version returns information. The second returns a recommendation with a stance, built for the real situation of the person asking. The second prompt is not longer for the sake of it. Each line answers something a real advisor would ask you before answering.

A work example: commenting on a financial close

How it is usually written:

Attached is the June close. Analyze performance and give me the main findings with a recommendation.

That returns a description of the file: correct figures and zero judgment. It does not separate what matters from what is incidental, and it does not say what to watch. With RICO:

R: Senior financial analyst, 15 years explaining closes to management. Explain clearly without losing rigor.
I: Analyze the June close, actual against budget, and build an executive report for the month.
C: Retail chain. Use only the data in the file. I care about the why and what I should watch.
O: Executive summary, KPIs with variance, three risks and three recommendations. One page maximum, board tone.

Two lines do more work than they seem to. "One page maximum" prevents the eight-page report nobody reads. "Use only the data in the file" is not decoration: without that line, the AI fills the gaps with reasonable assumptions that read exactly like data.

Why it works

Anthropic suggests a simple test in its prompting best practices guide: show your prompt to a colleague who knows little about the task and ask them to follow it. If they would get confused, the model will too. RICO is an orderly way of not skipping what that colleague would ask you.

Structure decides what the answer is good for. A vague prompt produces information, and a RICO prompt produces something you can decide on or present. That holds even as models improve.

When RICO is enough

RICO is still the default formula for daily work: an email to a supplier, meeting minutes, the reply to a lead. Newer models reason better, and that changes how much of the method you have to dictate. There is a second formula for tasks where a mistake is expensive. It is called STEER and it has its own guide.

RICO and custom instructions

You write RICO in every prompt, and it frames one specific task. Custom instructions, built with the 4Ps (Profile, Persona, Protocol and Presentation), are set up once and the tool carries them into every conversation. They work together: the 4Ps lay the base, and RICO sharpens each task.

Try it on one of your own tasks

Pick a task from this week where the AI gave you generic information.
Rewrite it in four lines, one per component. Start with the context, which is what people forget most.
Compare the two answers. If the second one helps you decide or present and the first does not, you have understood the method.