The RISEN Framework: Steps and Narrowing
RISEN stands for Role, Instructions, Steps, End-goal, Narrowing. Steps forces an ordered, verifiable plan and Narrowing forces at least one explicit exclusion — the two slots no other common framework has, and the two that stop a model from drifting on work with stacked constraints.
RISEN is the framework you reach for when a task has enough moving parts that the model can be right about each piece and wrong about the whole. This guide covers what the five slots hold, why the two unusual ones carry the framework, and the specific point at which their verbosity stops paying for itself.
By Andrei Bădulescu, Founder at VantagePrompt·Updated
There is a class of task where a model produces a fluent, plausible, subtly useless answer: it silently drops the third requirement, invents a value it was supposed to flag as missing, or solves a slightly easier version of the problem. That failure is what RISEN is built against, and two of its slots exist for nothing else.
What does RISEN stand for?
| Letter | Slot | What it holds |
|---|---|---|
| R | Role | The persona, in one sentence — same job it does in RTF. |
| I | Instructions | The primary directive in plain language, before any decomposition. |
| S | Steps | Ordered, concrete, individually verifiable actions. |
| E | End-goal | The observable outcome that means the work is finished. |
| N | Narrowing | What is out of scope, what to ignore, what must never happen. |
What makes a Step verifiable rather than descriptive?
A verifiable step has an outcome you could check without judgement. "Normalize phone numbers to E.164" either happened or did not. "Handle the phone numbers appropriately" describes an intention, and an intention cannot be skipped detectably — which is exactly how a requirement goes missing.
Order matters as much as content. Steps that depend on each other must appear in dependency order, because a model reading a list will generally execute it as written. Deduplicating before validating produces a different, wronger result than validating before deduplicating, and nothing in the prompt will tell you which one you got.
Steps (weak): Clean up the data, handle duplicates, and report problems.
Steps (strong): 1) Detect and report malformed rows — do not modify them.
2) Normalize phone numbers to E.164.
3) Deduplicate by email, keeping the most recently updated row.Why does Narrowing matter more than it looks?
Because every other slot tells the model what to produce, and a model with no stated boundary will produce more. Narrowing is the only slot in common prompt frameworks whose content is negative: what to leave out, what not to touch, which trade-off to honour when two requirements collide.
It is also where hallucination is cheapest to prevent. "Never guess a missing email" costs six words and closes the single most likely failure of a data-cleaning task. Discovering the same failure by reading the output costs considerably more.
A RISEN prompt whose Narrowing slot is empty is a longer RTF prompt with extra headings. If you cannot name one thing the answer must not do, the task probably has not earned this framework.
What is End-goal for, if the Steps already say what to do?
Steps describe the route; End-goal describes arrival. They come apart more often than they should: a model can execute every step and still hand back something unusable, because "run these three transformations" and "produce a dataset someone can load" are not the same requirement.
The useful End-goal is observable. "A clean dataset plus a summary of every change made" can be held up against the output. "High-quality results" cannot, and a slot that cannot be checked is a slot that was not filled.
Role: You are a data engineer.
Instructions: Validate and normalize a messy CSV of customer records.
Steps: 1) Detect and report malformed rows.
2) Normalize phone numbers to E.164.
3) Deduplicate by email.
End-goal: A clean dataset plus a summary of every change made.
Narrowing: Do not drop rows silently; never guess missing emails.When is RISEN the wrong choice?
When the task is one action with a known output shape. RISEN on a rename request produces a Role restating the obvious, an Instructions section paraphrasing the Task, a single Step, an End-goal restating that Step, and an empty Narrowing — five slots of ceremony around one sentence of ask, with the actual request now the least prominent thing in the prompt.
It is also the wrong choice for writing. The framework has no Audience and no Tone slot, and adding steps to a piece of copy does not make it more persuasive — that is CO-STAR territory.
When does VantagePrompt pick RISEN for me?
Automatically on the Reasoning and Structured use cases — both are decomposition work by definition. On Coding, and on Auto, it is selected when the request passes a complexity test: you stated explicit success criteria, or you stacked at least two hard constraints. Either signal means there is something for Steps and Narrowing to hold.
That test is worth reading in the other direction. If you want RISEN and are getting RTF, the missing ingredient is not length — it is constraints. Say what done looks like, or name two things the answer must respect, and the request now carries the evidence the selection rule is looking for.
When RISEN is selected, the optimizer makes Steps the longest section of the generated prompt and requires Narrowing to name at least one constraint or exclusion. That is the mechanical version of the same discipline: the two slots that justify the framework are the two it refuses to leave thin.
Frequently asked questions
- What does RISEN stand for in prompt engineering?
- Role, Instructions, Steps, End-goal, Narrowing. Role and Instructions set up the work, Steps is an ordered and individually verifiable plan, End-goal is the observable outcome that means it is finished, and Narrowing states what is out of scope.
- What is the difference between RISEN and chain-of-thought prompting?
- RISEN is a structure you write; chain-of-thought is reasoning the model performs. RISEN tells it which steps to execute and in what order. Chain-of-thought lets it work out steps you did not specify. They compose — a RISEN prompt on a reasoning task benefits from both.
- How many steps should the Steps slot have?
- As many as the task genuinely has dependencies, in dependency order. One step means the task did not need RISEN. Steps that could run in any order are usually a list of requirements wearing a plan's clothes — those belong in Instructions or Narrowing.
- What goes in Narrowing if nothing is out of scope?
- Then the task has not earned RISEN. A framework whose distinguishing slot is empty is a more verbose version of a simpler one — use RTF instead.
- Why does my complex-looking prompt still get RTF?
- Because complexity is measured by constraints, not by length. Selection looks for explicit success criteria or at least two hard constraints. Stating what done looks like, or naming two things the answer must respect, is what moves it to RISEN.
Sources
Put it into practice.
Run this technique in the optimizer.