The AI RICE Prioritisation skill for Claude

RICE provides a common set of questions for a roadmap discussion: how many people are reached, the estimated effect per person, confidence in those estimates and the effort required. It can make disagreements and assumptions more explicit. The resulting number does not remove personal judgement, stakeholder interests or the need to review evidence.

It is written for anyone who has to choose between too many options: product teams filling a roadmap, marketers comparing campaigns and channels, founders who want everything at once, and teams who need to order a content calendar or a list of operational projects. It is part of the skill library we make available for free from the AI and automation service, with no account and no sales email afterwards.

One thing you should know beforehand: RICE makes your assumptions explicit, it does not replace them. If you guess at Reach and wish for your Impact, out comes a number exactly as soft as that guess. That is why the strictest part of the file is not the formula but the bias-check protocol that runs, mandatorily, after every scoring round. More on that further on.

The formula is compact: (Reach x Impact x Confidence) / Effort. The numerator rewards ideas that reach many people, have a big effect per person and rest on evidence. The denominator punishes ideas that eat up a lot of capacity. The higher the outcome, the higher the idea sits on the list. What makes the formula clever is the role of Confidence: holding other factors fixed, a Confidence multiplier of 0.5 halves a score relative to 1.0; that does not by itself determine its rank. And dividing by Effort makes visible what teams systematically miss: that a small idea with an average effect often delivers more per month of work than a big idea with a massive effect.

The difference with choosing by gut feeling is shown fastest as a diff. On the left the sentences that usually settle roadmap discussions, on the right what the skill puts there instead.

The figures in this block come from the example scenario in the skill file itself. They are illustrations of the calculation method, not measurements from us or from a client.

Next come the four factor instructions you saw in FIG.02, each with its own probing rules and its own scale. And then comes the part that determines the output: a mandatory structure of five parts. First a short setup with the period, the main goal and the assumptions behind the scale. Then per item a RICE block with the number per factor plus a one-sentence justification, and for Confidence explicitly the weakest assumption.

Then the prioritised table, sorted descending, with per row the rank, item, the four factors and the score. Then a top 3 with, per item, why it is at the top, which assumption most urgently needs validating, and what you cut or postpone based on that top 3. And to close, the bias check, more on that below.

The file contains one worked arithmetic example for a four-item SaaS roadmap and a second marketing-calendar input with five ideas to work through. It does not contain a fully scored second example. Neither scenario establishes which kind of idea usually wins or validates a real business ranking.

The file closes with a list of five things the skill never does, a source credit to the original Intercom publication by Sean McBride, and five style rules. The most important of those: always answer in the language of the input, always show the calculation and not just the final score, and probe further the moment a factor cannot be determined from the input. If you want to learn how to build an instruction file like this yourself, the full explanation is at writing a SKILL.md NL.

Score inflation. If every item gets an Impact of 3, check calibration and evidence. The pattern prompts a review; it does not prove that the estimates are wrong.

Effort underestimation. If every Effort sits below 1, check omitted work and uncertainty. Add a justified buffer if needed; genuinely small projects do not prove underestimation.

Confidence guesswork. If more than half the items sit below 80 per cent, the first action is not building but research. The ranking is then not telling you what to build, but what to validate.

Reach inflation. Counting the same people more than once inflates the numerator. Use an agreed compatible definition across items. For unique-user Reach, deduplicate across the entire period; three monthly unique counts need not sum to quarterly unique users.

Incomparable periods. Placing an item measured per month next to one measured per quarter makes the figures misleading. Everything is measured over the same period, or it is not compared.

Alongside these five checks, the output itself flags three more things per round: which items sit below 70 per cent Confidence and therefore need validation first, which raw estimates were below 0.5 before applying the floor and which floor-scored items need a scope check, and which Reach numbers cannot be backed up with data. That way you never get just a ranking, but also the list of weak spots inside it. Anyone who wants to learn to test their own assumptions more broadly finds the wider thinking frameworks in the knowledge base.

After that, using it is simple: paste your list, name your main goal and your period, and ask for a RICE round. If anything is missing, the skill asks rather than guesses. If you get stuck anywhere while installing, every step is explained in the installing Claude skills NL guide.

The file is equally honest about when to leave it alone, and there are five such cases. With one or two options RICE is overkill: the scoring round is then more work than the decision. For pure urgency, where deadlines set the order, the Eisenhower matrix is the right tool.

For an existential go-or-no-go decision on a single project, a pre-mortem or an inversion is the right fit, because there you want to see risks, not a ranking. If items compete for the same resource, you need a capacity plan alongside the scores. And for strategic choices with a horizon of three years or more, the file points to frameworks such as Three Horizons or Wardley Mapping: RICE is a tool for the coming quarter, not for the coming three years.

The most important limit has already been mentioned, but it deserves repeating: the score is only as good as the input. RICE makes your assumptions comparable and exposes where they are weak, but it does not replace market research or a conversation with a customer. The file solves that by turning low Confidence into a validation task, so the weakest assumption automatically ends up at the top of your to-do list.

A good ranking is half the job: after that the top item still has to be built, written or launched. If an SEO or visibility project sits at the top of your list, the free SEO scan shows in a few seconds where your site stands right now, so your Reach estimate immediately rests on something. And if you want to talk further about what else AI can do for your business, from standalone skills to full automation, we are happy to do that in a conversation.