STEER: give the destination, let the AI choose the route

RICO is still the starting point for daily work. What changed with newer models is how much of that prompt you no longer need to write.

There are two situations, not two eras. In one, you know how the task is done and the AI carries it out. In the other, you can recognize a good result, but the route is not obvious even to you, and that is where the AI should choose it. For that second case I use STEER: you set the stakes and the destination, and the model picks the route. The name comes from the Spanish version, NORTE, and the letters here stand for Stakes, Task, End state, Edges and Review.

Two situations, two formulas

	Daily work	Judgment work
Formula	RICO	STEER
What you bring	The method	The stakes, the edges and the review criteria
What the model brings	The execution	The route and the self-check

What changed with newer models

Reasoning models, the ones that think before they answer, no longer need the step-by-step. Often the route they choose is better than the one a person would have written. If you hand them the procedure, you limit them. If you hand them the destination and the criteria, they choose the how.

The ones that also take actions (search, read files, create documents, send messages) add another demand: say what they can do on their own, what needs your permission and when to stop.

The providers' guides point the same way. Anthropic prefers general instructions over prescribed steps, and asks you to explain the reason behind what you ask and to state the limits. OpenAI recommends describing the destination instead of each step, and reserving ALWAYS and NEVER for rules that are truly non-negotiable. Google shares several principles, although its templates still include explicit steps. In short: two of the three agree clearly, and the third is halfway there.

The five blocks

Letter	Block	What it is	What to write	Example
S	Stakes	The why	What decision will be made with this, who makes it and what is at risk	"I'm presenting this to management and they will decide based on what I bring."
T	Task	What to solve and with what	The problem, not the procedure. Plus the material: which file, which sources and what counts as enough evidence	"From the attached report, tell me what really matters. Work only from the report; if something needs data that isn't there, flag it."
E	End state	What "finished" looks like	The shape and the quality bar, not just the format	"Three topics maximum. Finished means someone can decide by reading only this."
E	Edges	The guardrails, in three parts	What it does not do, what needs your permission and when it stops. Written as decision rules, not prohibitions	"If you're torn between two options, keep the one with more impact and name the other in one line."
R	Review	The mirror	What it checks before answering, against which criterion and what it must label	"Check that every 'why now' is concrete. Mark what you inferred and what is stated explicitly."

What does not appear in the structure is the steps. The model supplies those.

Edges in three parts

This is the block that changes most between a chat conversation and a tool that takes actions.

Part	In chat	In tools that take actions
What it does not do	Almost always enough	Necessary
What needs permission	Rarely applies	Critical: sending, deleting, publishing, buying, touching shared systems
When it stops	Implicit in the End state	Explicit: which condition marks it as done, or as blocked

Write the edges as decision rules, not prohibitions. Instead of "Don't list all 10 topics," write "Pick three at most and tell me which ones you would bet on." The first sentence only closes a door. The second says what to do when a choice has to be made.

Reserve ALWAYS and NEVER for what is truly non-negotiable: confidential data, irreversible actions. If everything is ALWAYS, nothing is.

The same task, written two ways

The task: prepare a synthesis of a report for a leadership meeting.

With RICO:

R: You are a senior strategy analyst with experience preparing material for executive committees.
I: Read the attached report and identify the 5 most relevant topics. For each one: description, impact and a recommendation.
C: Leadership will use it to decide the focus for next quarter. They don't have time to read the full report. They prioritize impact over completeness.
O: Table: topic, impact (high/medium/low), recommendation. One page maximum. Executive tone.

You give it the process and the shape. For a daily task, that is exactly what you want. With STEER:

[STAKES] I'm preparing next quarter's leadership meeting. They will decide which 2 or 3 topics get focus and budget, and they will decide based on what I bring.
[TASK] From the attached report, identify which topics really matter and which ones you would bet on. Work only from the report; if a conclusion needs data that isn't there, flag it.
[END STATE] Three topics maximum. For each: one line on what it is, one on why it matters now, and your recommendation. Finished means someone in leadership can decide by reading only this.
[EDGES] Take a position: if you're torn between two topics, keep the one with more impact and name the other in one line. If something depends on a decision of mine, mark it as [DECISION: ...] instead of resolving it.
[REVIEW] Before you give me the answer, check that every "why now" is concrete and tied to the report; rewrite the generic ones. Mark what you inferred and what is stated explicitly.

The difference is in the verb. RICO asks for "the 5 most relevant topics": count and describe. STEER asks "which ones you would bet on": choose and defend. The first produces a summary and the second a position.

An example with actions: the client inbox

Imagine a tool with access to your email and your calendar. Here the edges stop being a detail:

[STAKES] On Friday I need to arrive with every client answered before month-end close; I don't want anyone waiting over the weekend.
[TASK] Go through this week's client emails that have no reply from me and prepare a response for each one. Use each thread's history as context.
[END STATE] One draft per email, in my tone, and a final list with: client, topic, and whether the draft needs anything from me before it is sent.
[EDGES] You can read emails and create drafts without asking. Before sending, deleting or scheduling anything, ask me for confirmation. Stop when every pending email has a draft or a note on why it doesn't.
[REVIEW] Before you give me the list, check that no draft promises dates, prices or commitments that are not in the thread. If one does, flag it.

All three parts of the edges are there: what it does alone (read and create drafts), what needs your permission (send, delete, schedule) and when it stops (when every email has a draft or a note).

From RICO to STEER

RICO	STEER	What happens
Role	Optional, at the start	Loses weight, doesn't disappear. You can put a role line first when the domain calls for it
Instruction	Task	The central change: it goes from "do A, then B, then compare" to "this is the problem"
Context	Splits: the why goes to Stakes, the material goes to Task	Gains weight: it now includes which decision is made with this and on what evidence
Output	End state	Goes from format to a definition of finished
Doesn't exist	Edges	New. Once you drop the method, the guardrails keep the model from going where it shouldn't
Doesn't exist	Review	New. Separates a convincing text from a defensible analysis

STEER does not replace RICO. It collapses the prescriptive part of the Instruction and raises the value of Context and Output.

When to use which

Criterion	RICO	STEER
Type of model	Fast, everyday	Reasoning, high effort, or a tool that takes actions
Type of task	Narrow. You already know how it's done	Open. The model may find a better route
Do you know what a good result looks like?	Yes, and you know the route	You can recognize it, but the route isn't obvious even to you
Risk	Low: a draft, a summary, an email	High: it goes to a decision, to a room, or it takes actions
What you bring	The method	The judgment and the limits
Examples	Email to a supplier, meeting minutes, reply to a lead	A financial close before presenting it, strategic prioritization, an agent processing your inbox

Most daily work is still RICO. STEER is reserved for when judgment matters more than speed, or when the AI will act and not just answer.

Two traps STEER does not solve alone

The convincing output. With the most capable models, a vague prompt no longer produces obvious junk. It produces something that sounds good: well written, well structured and sure of itself. The failure happens when you accept it without checking. The AI does not know when a number is wrong, and it explains it just as well. The judgment is yours. That is why the Review block asks you to label what the AI inferred apart from what the material says.

Writing less without judgment. If you teach STEER as "now write shorter prompts," you remove the method and the criteria with it. The result looks obedient and has no judgment. Remove the method, not the criteria. OpenAI points in the same direction: when you simplify a prompt, keep the success criteria and the stop conditions, and before raising the model's effort, check whether the prompt is missing a criterion or a verification. If you can't say what "finished" looks like, RICO is still your floor.

Try it on one of your own tasks

Pick an open task from this week, one that ends in a decision. Avoid a routine email.
Write it in the five blocks, in that order. Start with the Stakes: which decision is made with this and who makes it.
Add at least one edge written as a decision rule.
Compare the answer with the one a RICO prompt would give you. The difference you are looking for is a position, not a summary.