Copy Lab: An AARRR-Driven Copy Inspiration Generator

Origin Story

Why I Built Copy Lab

This idea came to me as a side project while I was studying marketing theory in a systematic way. I have always felt that AI cannot really finish a piece of marketing copy for you. Tone, restraint, and a real feel for the audience are still things people understand better.

But I was curious about a narrower question: can AI help at the intersection of inspiration and consumer psychology? By mixing theories from persuasion psychology and consumer behavior with a real product scenario, it could offer an angle that has real theoretical grounding, one that still needs a person to shape it further.

Building it did two things at once. It forced me to actually use the theory I was learning and to explain clearly why each messaging angle holds up. It also let me test the limits of AI: whether it could reliably tell an external trend apart from an internal execution tactic, and whether it could stay consistent about following the writing rules.

1. Positioning

Copy Lab is an AI copy inspiration tool for growth marketers. Based on a product's lifecycle stage and a specific scenario, it draws on consumer psychology to generate three inspiration packages at once, each built around a different angle. The goal is to help marketers find direction fast, not to produce finished copy.

2. User Inputs

2.1 Basic Inputs

Product Flow

Figure 1. The six-step input flow. At step 4, choosing a Moment, Activation, Retention, and Revenue each split into two groups: Momentum and Recall.

2.2 Moment

Most Moments under each stage capture a point where the user just did something right (Momentum). Activation, Retention, and Revenue each add two Moments that capture a point where the user is starting to drift (Recall). Acquisition users have not touched the product yet, and a Referral share is a one-off event, so neither stage has an ongoing drift state to observe, and neither one gets a Recall group.

A small amber dot marks Recall. Everything else is Momentum.

See 3.2 for how Moment type shapes the writing.

2.3 Channel Matrix

There are four channel configurations, set by stage and product type together. A Recall Moment uses the same channel configuration as its stage, so it needs no extra logic.

3. AI Generation Logic

Each API call generates three Trend Packages. The front end lets people flip through them with a Shuffle button, without calling the API again. A single question rarely produces the best possible idea on the first try, so Copy Lab generates three different angles at once and lets the marketer compare and choose, rather than betting that one generation is already the best answer.

3.1 Output Overview

See 3.2 through 3.6 for the details and constraints.

3.2 trend_text: The Decision and How to Write It

The JSON field is always named trend_text. The rules and examples below go straight into the system prompt, so they work as both an explanation for you and the actual rules that constrain what the AI writes. Its meaning is set by a two-step decision:

Figure 2. The decision logic for trend_text. First check whether the stage is Acquisition. If yes, it is a Trend Pulse. If not, check the Moment type next, which splits into Momentum Idea or Recall Idea.

1. Acquisition to Trend Pulse

The AI should describe an external, observable trend. It can come from a shift in consumer behavior, content culture, or platform preference, and it should explain why this trend makes the current angle work. It must not use specific numbers unless they are public, verifiable industry figures, and it must not dress up internal data or an execution tactic as an external trend.

2. Not Acquisition, Momentum, to Momentum Idea

The AI should anchor on the psychological state of a user who is currently carrying momentum forward, and explain why this messaging angle will resonate. It should not generalize into something like "users typically feel...". Keep it to two or three sentences, with the last one landing on the copy strategy.

3. Not Acquisition, Recall, to Recall Idea

The AI should acknowledge that some time has passed, but stay neutral in tone and avoid asking why. It should land on how easy it is to pick back up, not on what the user has already invested. Guilt, blame, or overly emotional language is not allowed. Do not write things like "we miss you" or "why did you stop".

Moment type also shapes the angle recommendations in 3.3 and the Recall-only tone rule in 3.6.

3.3 The Messaging Mashup

This is the Messaging Angle. Each of the 3 Packages generated must use a different angle. All seven angles are backed by psychology or consumer behavior research; the full write-up is in the appendix at the end.

What the stage-to-angle guidance means: the table below is not a hard rule, it is the AI's default preference. At Acquisition, for instance, the AI normally picks from Social proof, Curiosity or contrast, and Identity or aspiration first, but it can choose a different angle if that fits the product better this time. The only hard requirement is that the 3 Packages cannot repeat an angle. This keeps the angle choice grounded and matched to the stage's psychology most of the time, while still leaving room for the AI to use its own judgment instead of being locked into a fixed set.

The three Recall rows are combined into one because the psychological state driving them is really the same thing in each case: attention has drifted, and the person needs an easy, low-pressure reason to start again. That does not change by stage.

3.4 A/B Test

Why is it designed this way? Copy Lab already generates 3 Packages with different angles on every run. Instead of inventing a comparison group that the user can never see, it makes more sense to let these 3 Packages serve as each other's reference point. The comparison is real, and the user can flip over and look at it directly instead of taking a claim on faith.

What is it? Each of the 3 Packages carries its own ab_plan. This is not about pulling 2 of the 3 out to run a separate comparison; every Package comes with its own suggestion for which of the other two would make a good reference point if you wanted to test this angle. One generation produces 3 separate ab_plans, and each one points at one of the other two Packages.

How does it work? The model generates all 3 Packages in the same call, so when it writes the ab_plan for Package N, it already knows what angle the other two are using. It picks whichever of the other two is the most interesting comparison right now and writes that into the comparison field. This is not a fixed pairing like Package 1 against Package 2 and Package 2 against Package 3; each Package makes its own judgment about which comparison is most worth making.

Is this a strict A/B test? A rigorous A/B test usually needs to satisfy three things: it changes only one variable while holding everything else constant, it locks in the metric to watch and what happens after a win or a loss before the test starts, and it has random assignment and enough sample size to separate a real difference from noise. Here is an honest comparison against how Copy Lab is designed:

Example. Using BeautyPlus and Acquisition again, if the 3 Packages generated each use a different angle, their comparison fields might point at each other like this:

This is just one possible pattern, not a fixed loop of 1 to 2 to 3 to 1. The actual choice is up to the model's judgment about which comparison is more interesting.

JSON structure:

A full ab_plan example for Package 1:

3.5 Callout

Why does this field exist? Copy Lab is positioned as a starting point for ideas, not as finished copy. The point of Callout is to hand over one more execution idea, beyond the copy itself, that the user probably would not have thought of on their own.

The rule: The AI picks whichever of the four directions below fits the current package best and writes just that one; it does not need to cover all four. It must not touch send timing, audience segmentation, or channel sequencing, and it must not suggest product features either. Those are areas marketers already know better than this tool does, and it should not overstep into them.

The payoff: For the user, this is the single most memorable part of the output, and it raises the odds they come back or share it. For the product, it cheaply separates an AI copy tool from an AI inspiration tool, which reinforces what Copy Lab is meant to be.

3.6 AI Writing Constraints

The sections above cover what to do. This one covers the hard limits on what not to do, plus the format limits for each channel.

• Never invent specific numbers, such as user counts, percentages, or days, unless they come directly from the selected Moment (for example, "Came Back Day 7" can mention "7 days")
• Once an angle is set for a Package, it cannot switch midway or introduce a new framing
• Never name the selected Moment directly in the copy (do not write "you just finished the tutorial")
• Never generate fields for a channel that was not selected
• A Recall Moment must never use a guilty, blaming, or overly emotional tone
• If the AI cannot recognize the product name because it looks made up, it must not invent a category or positioning and force a generation. It should show "This product could not be recognized, please check the name" and stop

Format limits by channel

4. Frontend Interaction

4.1 Form Flow

4.2 Result Layout

Shuffle logic. It switches between the 3 already-generated Packages without calling the API again. Copy All logic. It copies the entire content of the currently shown Trend Package to the clipboard.

Appendix: Theoretical Grounding

Appendix: Theory

AARRR: A Practical Framework, Not an Academic Theory

AARRR (Acquisition, Activation, Retention, Referral, Revenue) was first shared publicly by Dave McClure in 2007, in a talk called Startup Metrics for Pirates. He kept refining and re-presenting the framework over the following decade. The original material can be found on his SlideShare account (slideshare.net/dmc500hats) and his blog, Master of 500 Hats (500hats.com). McClure is a founding partner of the venture capital firm 500 Startups.

To be honest about it: AARRR is a practical metrics framework built for startups, not a peer-reviewed academic paper, and there is no original research paper for "AARRR itself" to cite. It is better understood as a simplified, easy-to-remember, practical version of the larger idea of customer lifecycle management.

Related Academic Framework: Lemon & Verhoef (2016)

The Theory Behind the 7 Messaging Angles

Tap a card to open it.