What I took from it
The parts of this that feel like design and the parts that feel like prompt engineering aren't separable. Deciding what counts as good tone, what has to be asked before generating, where the source of truth lives, and how the cascade to other surfaces works, those are design decisions expressed in prompts, project structure, and skill scoping.
The closer I get to the model, the more the work looks like designing a system, and the less it looks like designing a screen. The screen is downstream of the system.
Read the full case study
The problem
Some Bibel TV fundraising letters raised noticeably more than others, and nobody could say exactly why beyond a guess about the season. Every campaign still needed a letter that read warm, specific and unmistakably Bibel TV — drafted from scratch, several times a year.
The follow-up content built to carry the same message into the app, the web and video scripts had two options: get copy-pasted, and lose the specificity that made the original work, or get written from a blank page again, and lose the time. The blank page kept eating the calendar.
What I built
1. A performance analysis of the source material
I went back through the fundraising letters and their results, identifying what separates the over-performers from the under-performers. Seasonality was one axis. Style, sentence length, warmth, how directly the ask was framed, and how much specific reference to Bibel TV programming was woven in — that was the other. That analysis is the ground truth the whole ecosystem sits on.
2. A Claude Project that guides the generation with follow-up questions
The analysis, together with the style guide, is encoded into a Claude Project. Crucially, the Project doesn't just take a prompt and produce a letter. It asks follow-up questions before generating, clarifying the campaign occasion, the audience segment, the specific programme or milestone the letter should reference. The interaction is a multi-turn clarification loop with a human, not a single-shot completion.
3. A web prototype for the generation flow
The flow isn't chat-only. I built a web prototype that gives the generation a proper surface: inputs for the context Claude needs, a side-by-side view of the working draft, iteration in place. Prompt engineering shows up in the model's behaviour; the prototype shows up in the human's ability to drive it.
4. A Claude skill for multi-surface follow-up content
The fundraising letter is only the first surface. The same message has to land as an in-app notification when a viewer opens the Bibel TV app, as a banner on the web, and as a video script for the on-air fundraising spots. I built a Claude skill that takes the letter's tone and message as its ground state and produces follow-up content across all three surfaces. One voice, three surfaces, one source of truth.
How the four pieces fit together
Analyse the source material
Read past letters against their performance; identify seasonality and style patterns; write the findings into a style guide the model can consume as context.
Encode into a Claude Project
Style guide + analysis loaded as project knowledge. Project instructions ask follow-up questions before generating: what campaign, what audience, what programme is anchoring this.
Draft in the web prototype
The prototype gives the flow a proper working surface. Enter the context, iterate on the draft, keep control over what the letter actually says.
Cascade with the Claude skill
Once the letter is decided, the skill produces the in-app notification, web banner, and video-script versions. Same voice, adapted to surface.
The design decisions that mattered
Ground the model in analysis, not in vibes
The Claude Project's job isn't to imitate a generic fundraising voice. It's to encode what actually works for Bibel TV, the seasonality patterns, the specific stylistic choices that separated the over-performers from the rest. Prompt engineering here is really source-material engineering: the leverage is in the analysis, not the phrasing.
Ask follow-up questions before generating
The most common AI-writing failure mode is a plausible-sounding first draft that misses the specifics only a human can supply. Building the follow-up questions into the Project structure turns that failure mode into a design choice: the AI won't generate until the operator has answered what campaign and what programme this is for.
Multi-surface cascade with one source of truth
The skill treats the letter as the source and the notification / banner / video script as derivatives. If the letter changes, the follow-up content stays in sync. If Bibel TV's tone shifts, we change it in one place. That's a design decision about where the source of truth lives, before it's a prompt decision.
Solo end-to-end at the model-adjacent layer
I built all four artefacts: analysis, Claude Project, web prototype, Claude skill. That is the point of the story, not a team boast. This is the shape of the work when a design leader stops designing around AI and starts building at the prompt-and-context layer.
Curious why the fundraising AI won't generate until it knows what campaign and what programme this is for — and how one letter becomes an in-app notification, a web banner, and a video script without anyone rewriting it? Get in touch.