CopyCat: AI Copy That Partners Actually Understand

The problem: an app its users can't read

Porter's delivery partners live inside the partner app: orders, payments, penalties, support. But over 80% of them struggle with reading and comprehending the app's English interface. Many navigate by memorising the shapes of a few familiar words. When the copy gets abstract (support flows, task instructions, penalty explanations), comprehension collapses: wrong taps, stalled tasks, support tickets that a clearer sentence would have prevented.

The standard fix is a copywriting pass. But copy debt regenerates: every new feature ships new strings, written by whoever built it, in whatever tone they had that day. The problem isn't a batch of bad strings. It's that the org had no system for producing partner-legible language.

The reframe: don't rewrite the copy, systematise the voice

CopyCat is two layers, and the order matters.

Layer one: a persona-first tonal matrix. Before any AI, I built the framework: who is reading (new partner vs. experienced partner), in what moment (task-critical, money-related, support, celebration), and what tone each cell demands: direct instruction, reassurance, plain-language explanation. The matrix encodes the language partners actually use, the words observed from partner conversations, not translated corporate English.

Layer two: AI tooling that enforces the matrix. A designer or PM uploads the screen they're working on; CopyCat reads the context, detects the intended tone from the matrix, and generates partner-friendly copy options that fit it. The AI isn't freestyling. It's operating inside the framework, which is what keeps the output consistent no matter who's asking.

Early versions produced generic, could-be-any-app copy, the classic LLM failure. The fix was doubling down on layer one: richer screen context, tone constraints from the matrix, and continuous validation of outputs with the designers using it. Simplifying language turned out to be a deeply contextual exercise: not shortening text, but aligning it with the reader's reality.

Proving it

The sharpest external validation: CopyCat won Porter's company-wide AI hackathon, and from there moved into the org's actual workflow, with designers and PMs generating on-voice, partner-legible copy without waiting on a copywriter, and support/task flows getting rewritten through it. In our post-launch checks, comprehension accuracy on rewritten flows improved roughly 20%, and the design team's time-per-copy-rewrite dropped by about 40%.

What I carry forward

AI didn't solve the comprehension problem. The tonal matrix did. The AI made the matrix enforceable at scale. That ordering (framework first, model second) has become my default for every AI intervention since: the model is only as good as the system it's allowed to operate in.