The AI Hook Model Builder skill for Claude

The skill focuses on recurring use rather than pure acquisition or advertising. Users may fail to return for many reasons, including insufficient value, poor fit or unsuitable timing. A loop is one design hypothesis to test, not a universal solution to a forgotten product or a reason to assume that marketing cannot help.

Nir Eyal described a model for this in 2014 in Hooked: four stages that follow each other in a loop, trigger, action, variable reward and investment, as a design hypothesis that useful repeated experiences may encourage return visits. What you get here is not that general explanation, because there is a book for that. It fills the model in for this specific product, this audience and this core behaviour.

It is written for product managers, founders, designers and marketers working on retention, onboarding or activation. You do not need an existing product: with an idea, an audience and a core behaviour you can already work through the stages. What you do need is one action you want to see more often, not a collection. Core behaviour is singular, and that is the first place the skill corrects you.

And then the part that sets this skill apart from most retention advice you find online: the ethics check is a mandatory step, not a footnote. It tests against Eyal's own Manipulation Matrix, and the outcome is explicit. For anything that causes harm with heavy use, such as gambling or impulse credit, it does not build a loop at all. This skill belongs to our AI and automation service and sits alongside the rest of the skill library.

The trigger comes at the wrong moment or is not tied to anything. Explore whether the cue fits an existing user need and an appropriate moment. Boredom, unease or fear must not be invented or induced. A proposed connection needs validation; the file no longer predicts that an unsuitable notification is disabled within a week.

The threshold is too high for the motivation that is there. FIG.03 uses Fogg’s conceptual model to investigate motivation, ability and prompt. It does not calculate behaviour or establish that simplifying an action is always cheapest. Seven fields is an added fictional assumption in the example, not proof that ability is the only problem.

The reward does not vary. Distinguish a variable reward from predictable feedback. Predictable feedback can still be useful; reinforcement research does not prove its effect disappears in every product. For each proposed implementation, describe what varies and why that variation serves the task.

The user puts nothing into it. Data, preferences, content, contacts, structure or time may improve usefulness, but those effects and switching costs are hypotheses to test. Identify a possible next cue without trapping the user or assuming that investment always improves the product. An investment that does not load a next trigger does not close the loop, and then you do not have a loop but a straight line with an end.

In diff form: the patterns the skill refuses, next to what it puts in their place. This is not a style preference, it is a hard rule in the file.

After that comes a theory chapter that explicitly names four sources and links each to a stage. Eyal's own Hook Model, the Fogg Behavior Model at the action stage, Skinner at the variable reward, and the IKEA effect plus cognitive dissonance reduction at the investment. The ethics check also comes from Eyal, via the Manipulation Matrix. That is not decoration: because the reasoning sits in the file, Claude can explain, at every choice, why that choice is made rather than just what you have to do.

The core is a fixed method of seven steps, with an initial ethics gate after context and before design, and Step 7 repeated afterwards. Proposals and assumptions must be labelled. Step 1 is the context stage: product, audience, core behaviour and frequency. Core behaviour is one action and not a collection, and frequency is a cadence to assess for this product, not evidence by itself that a habit will form.

After context and the initial ethics assessment, Step 2 proposes external and internal triggers. Step 3 examines motivation, ability and prompt qualitatively, then three feasible simplifications. Step 4 chooses a primary reward type and three implementations with explicit variation. Step 5 tests whether user investment improves usefulness and can support a next cue; increased switching cost is not assumed.

Step 6 summarises the loop in five lines and proposes three two-week experiments. Each needs a hypothesis, comparison, measurement and decision rule, with unwanted effects considered. A noisy result may be inconclusive rather than refuting the hypothesis. Step 7 repeats the ethics assessment already performed before design.

Alongside the seven steps, the file contains an eight-point input checklist, with the instruction to only ask about what is missing, and the rule that a missing figure about current usage is noted as [TO BE COMPLETED] rather than measured itself. The output format has eight parts: context includes the initial ethics outcome, and the final item repeats the check on the proposed design. After that come the shortened example from FIG.01, seven things the skill never does, five honest limits of the framework, and a source list of seven titles.

If you want to write a file like this yourself, or adapt this one to your own product, that is exactly the point: it is released under the MIT licence. Writing a SKILL.md NL explains how to build a file like this, from frontmatter to refusals block.

Nir Eyal laid the model down. Hooked, written with Ryan Hoover and published in 2014, describes the four stages and the way they reinforce each other. Eyal also wrote Indistractable in 2019, and that book is not there by accident: it describes the other side, how users arm themselves against loops. Anyone who reads both designs more carefully.

The action stage comes from BJ Fogg. His Behavior Model, presented at the Persuasive conference in 2009, states that behaviour arises when motivation, ability and a prompt come together: the current shorthand B = MAP, a conceptual model rather than multiplication. Fogg also wrote Tiny Habits in 2019, and that is in the list for linking new behaviour to existing routines. That explains why the input checklist explicitly asks whether there is an existing routine the behaviour can latch onto.

Reinforcement research includes Ferster and Skinner. Their 1957 Schedules of Reinforcement examines fixed and variable schedules. That experimental work does not establish that novelty guarantees attention or every predictable product reward stops working. The four-stage product model is not a measured law.

The investment leans on Norton, Mochon and Ariely. Their paper was available online in 2011 and published in the 2012 volume. Successful completion of the studied tasks increased valuation; that does not establish that every product investment increases value or switching cost. That is exactly where the ethical line sits too, because you can use that effect to make something better, or to trap someone.

The context comes from Wood and Neal. Their 2007 paper in Psychological Review describes habit formation as context driven behaviour: the environment sets the action in motion, not a conscious choice. That is why the skill asks for the exact moment and channel for every external trigger. A trigger without context is an idea, not a trigger.

Anyone who wants to apply this way of thinking more broadly than a product can put the Atomic Habits Coach skill alongside it for habits in people rather than in products, or the Fogg Behaviour Model Coach skill for the qualitative model from FIG.03 on its own, apart from the Hook Model. And for the persuasion side of the same question there is the Cialdini Persuasion skill.

After that, using it is simple: describe your product in one sentence, name your core behaviour and the frequency you want to see, and the skill works through the seven steps. It only asks about what is missing from the eight-point checklist. If you do not have a figure for current usage, that is not a problem: it marks the missing value. Essential benefit or harm information must be clarified before design. If you get stuck on the installation itself, the full route is in installing Claude skills NL. If you want to learn how to set this kind of work up with your team, that is the subject of our AI training.

There are also situations where you are better off leaving it alone, and the file names them itself. For pure acquisition and advertising questions this is the wrong model: that is about the first visit, and the Hook Model is about return visits. And for products that cause harm with heavy use, it refuses to serve.

Then the honest limits of the framework itself, because they are all in there. The model describes behaviour that recurs often, in a short cycle. For products you use twice a year, such as a tax return or a mortgage, it is not the right model. Those infrequent tasks are outside this skill’s intended short-cycle scope.

The evidence is partly practical observation and partly laboratory research. Variable reinforcement is well researched, but the four stages as a whole are an organising model and not a measured law. It is a good way to structure the conversation, not a law of nature you can invoke.

This skill requires a credible user-benefit hypothesis. Habits can also be unhelpful; habit formation does not establish product value. If you doubt whether the underlying usefulness is there, the work starts somewhere else: with testing the assumption itself, for instance with the Mom Test Coach skill for your customer conversations.

Measurements take weeks to months. A two-week experiment gives direction, not proof. And what reads as a friendly reminder in one country or segment feels pushy elsewhere. Test triggers with real users before rolling them out: a plan on paper does not establish that its timing is useful or welcome.

A good loop brings people back, but someone still has to come in first. If the inflow lags behind, the real work sits in front of this page: getting found by the people looking for you, in Google and in AI answers. The free SEO scan shows in a few seconds where your site stands. And if you want to talk through what else AI can do for your business, from individual skills to full automation, we simply do that in a conversation.