Skip to main content
Back to the villageDONGGU’S FIELD NOTESCOPYKLE
The workshop · selected work
PLACE
스타트업
ROLE
Full-Stack
STARTED
2024-06
★ Case 02 / 5Forward Deployed Engineer Donghyun Kang

Copykle AI Marketing Solution

Translated the marketer's free-form one-liner into structured input that produces good copy and clears legal review.

스타트업Full-Stack2024.06 ~ 2025.06
01.Discover

"Free-form input" really meant uneven output.

The LLM in the marketer's hands took "write me ad copy" and returned something generic. Good output only came when the input carried target, tone, use case, and brand context — and writing all that out every time was not their actual job. The answer was not to teach marketers a new skill called "prompt engineering"; it was to have the system structure the input for them.

And raising the copy quality alone was not the end. An ad only becomes "usable copy" after it clears the ad-disclosure law and the review guidelines, and running that check by hand each time was unrealistic. The endpoint of structuring was not the input — it was the output verification.

Field notes

Note 01

The "free-form" input was not actually free

A line like "write me ad copy" returned generic output every time.

Note 02

The context was missing from the input

Of the four axes — target, tone, use case, brand — typically only one or two arrived filled in.

Note 03

Prompt engineering was not the marketer's job

Forcing marketers to learn a new skill was not going to land.

Note 04

Passing as copy is not the same as passing as an ad

Someone was running the ad-disclosure-law and review-guideline check by hand for every piece of copy.

02.Build with AI

Let the backend write the prompt and run the legal check on the user's behalf.

Decision 01

Free-form → choice-based UI

BeforeA single free-text box: "write me ad copy"
AfterVetted choices — target, tone, use case, brand — branching the input

tradeoffLess expressive freedom vs. more consistent output — chose consistency

Decision 02

Bury prompt engineering in the backend

Take the user's option combo and have the system compose it into a vetted prompt template. The UI surfaces only 6–7 options; the structured prompt runs behind them. OpenAI, Claude, and others are routed by tone and use case, so the cost/speed/tone tradeoff is decided by the system.

Decision 03

RAG ad-review pipeline

BeforeA person cross-checking every piece of copy against the disclosure law and review guidelines
AfterAd-related law and review documents embedded into a vector DB; generated copy matched against the relevant clauses via RAG; compliant-or-not decision rendered (MVP validation stage)

tradeoff1:1 human review (accurate but slow with gaps) vs. automated RAG (faster, broader coverage, framed as a system assist)

Decision 04

WebFlux async + experiment-driven tuning

Spring Boot 3 + WebFlux absorbed LLM response latency through async. One request could call several tone variants in parallel without stacking up the wait. A resort ad-banner case ran tone, length, and keyword variable combinations to find the passing line, and that pattern was locked in as the system default.

Copykle main
03.Prove with metrics

The places the system took over from people.

Input shape

Free-form → choice-based

6–7 option branches

Prompt authoring

User → backend

Template composition + model routing

Legal check

Human → RAG

Disclosure law and review guidelines embedded

Tuning method

Guessing → locked by experiment

Resort-banner case

What this case is really about
“A full-stack engineer who translates free-form prompts into structured inputs so marketers never have to see the AI under the hood.”
Discover

The real load on the marketer was not the copy itself, it was getting the ad through legal review.

Build

Built a choice-based UI, server-side prompt composition, and a RAG check for ad-review compliance.

Prove

Shipped the MVP, locked the passing pattern through a resort-banner experiment, and moved the manual legal check inside the system.

Linked principlesStandardizationOwnership