Branding and Context Building: Very Nerdy Stuff

As I’m writing this, we’re hot off the launch of the new Dragonfly brand and website, which debuted at their 10th anniversary celebration last night. 

The client held a party in the event space at the back of Plenty Mercantile, gathering friends and supporters to toast how far they’ve come since their humble beginnings. I wasn’t able to attend (dad-duty calls), but by all accounts, the new brand was received well. Melissa and Wendy both cried during the brand launch video. We don’t have an exact count on the total number of criers, but I’m confident we hit at least five.

I’m in a retrospective mood. The Dragonfly website was my first project where I leaned heavily on AI to compress the timeline. Now that it’s live, I want to walk through the process — mentally retrace my steps, look at how things started, what we scoped, how long we projected each phase to take. 

I want to notice where things went well and where they got frustrating. 

I want to walk through the individual conversations I had with Claude and recap what the actual doing-of-the-work looked like — because by the time I finished, so much had happened that I didn’t have a super-clear memory of all the myriad tasks and miniature side-quests I’d run with it.

Stumbling My Way into Context-Building

After doing so many projects with Kiowa Casino’s email campaigns, I’d picked up some tips and tricks for getting consistently on-brand output from ChatGPT. The core lesson: the more background information I provided, the better the output. “Better” is a bit subjective here — it could mean more accurate, more on-brand in terms of voice, or just more varied options for headlines. The model had preexisting knowledge it could draw from and reuse.

For example, my ChatGPT instance knew that Kiowa had different types of emails — promotions, events, Sunday free gift, VIP free gift, tournaments — and each had its own format. VIP emails were written to sound like a personal note from GM Darius McGee, complete with a signature. Promotions followed their own structure: a thematic intro paragraph, chunked-out offer details, and a closing paragraph inviting guests to visit. After feeding the model dozens of examples of existing emails, it got a pretty good feel for what was needed. 

It was only much later that I could give this process a name: context-building. I’d arrived at the technique intuitively, without a word for it.

For Dragonfly, I built that initial context in layers. First, I uploaded the website scope of work document so it understood the project’s contours — the timeline, action steps, goals, requirements. A kind of “meta” context for the project at large. 

Then I uploaded a PDF of the brand strategy, which recapped the discovery process and made recommendations for the brand’s direction. I also uploaded a PDF export of the visual and verbal brand presentation — the one the client saw in the room — which contained information about brand personality, archetypes, voice and tone, mission, vision, values, and whatever other language we’d used to present the work in its best light. It was, after all, a presentation designed to convince, not a systematic style reference. But it was a start.

With that initial base of knowledge in place, I had enough basic context to get started in the right direction.

The Phases

Phase 1 — Discovery & Architecture (Early March)

I used Claude mostly for gathering facts and information. To get the lay of the land — figure out the architecture of their current site, take an inventory of pages, decide what to keep or cut. We worked through this iteratively until we arrived at a revised, finalized scope of work. 

• Existing site crawled via sitemap and 9 core content pages synthesized into a Content Intelligence Brief
• New site architecture defined: reduced from 30+ pages to 5 core pages + dynamic sections
• Scope of work document built collaboratively
• Interactive React sitemap diagram produced and added to scope doc
• PDF export of sitemap produced for client presentation

Phase 2 — Keyword Strategy & Content Planning (Late March)

SEO has been a big part of my world this past year, and this project was the perfect start-to-finish opportunity to apply what I’d learned. 

I started with keyword research in SEMRush, using “human trafficking” as a seed keyword and seeing where it led. I loaded a batch of keywords into a tracking list, which gave me good decision-making data on each one — keyword difficulty, search volume, intent, the works. 

I called this my “keyword bible,” because it became my canonical source of truth for all keyword decisions. Having it as a CSV mattered too — it’s a clean, machine-readable format that plays nicely with Claude.

• SEMRush MCP integration connected; 303-keyword position tracking list pulled from ‘human trafficking support okc’ campaign
• All 303 keywords triaged into three buckets: core page assignments, blog candidates, and discards
• Keyword bible CSV built and uploaded as the authoritative project file
• Blog confirmed as the destination for informational and long-tail keywords that don’t fit core pages
• Primary keyword selected for each core page based on intent, volume, and keyword difficulty

Phase 3 — Content Briefs (Late March)

• About page content brief written first, used as the format template for subsequent briefs
• Get Involved page brief written, anchored by ‘how to fight human trafficking’ (KD 50)
• Donate page brief written, anchored by ‘donate human trafficking’ (KD 37, ranking #6 at time of brief)
• Copy prompt engineered and refined: Content Brief Page Outline made the authoritative source for section structure, making the prompt reusable across all pages

Phase 4 — Copy Drafts (Late March through April)

This is where we started generating actual copy drafts. The idea was to have Claude produce page copy as raw material — something to mold and refine. It might not say exactly the right thing with the judiciousness of a seasoned copywriter, but it’s a structured starting point that follows an agreed-upon flow, chunk size, and hierarchy.

Pretty quickly I found myself in a nice sequential groove:

1. Create a content brief based on the keyword bible
2. Generate a good-enough first draft based on the content brief
3. Send the draft to creative

Things were going well. Then I tried to use Claude to generate layouts — and the hiccup wasn’t quite what you’d expect.

The first issue was incompatibility. Claude is great at generating code, but as of this writing, it doesn’t output code that integrates natively with Elementor. On paper it seems like it should work — Claude can output JSON, which is the format Elementor uses for importing and exporting templates. But in practice, it produces old-school Elementor output, pre-flexbox, with workarounds like inline styles inside text boxes to fake cards. It technically works, but you’re essentially rebuilding it in Elementor anyway.

The second issue was predictability. Claude’s layouts were clockable. Similar to the telltale em dash, or the “it’s not X, it’s Y” phrasing quirk that LLMs have adopted, this habit shows up visually too — lots of little decorative details sprinkled throughout, very predictable placement of elements. You can spot it.

The upside is that going from a flat text file to a rough visual layout is fast, and it does make it easier to see how copy will flow in a web context. We all know that copy that reads well as plain text doesn’t always land the same way on a page.

After several hours of trial and error — trying to force the Claude-to-Elementor pipeline to work — I gave up and built the site directly in Elementor, using the Claude mockups as visual references. Having been through it once, I know this is the right call. It might be worth trying Elementor’s own AI layout generation at some point (it’s been a feature for at least a year), but I wasn’t ready to spend money on credits for this one.

Phases 5 & 6 — Build & Polish

• Quick Exit button implemented in minutes with Claude’s help
• Blog category naming conventions worked through collaboratively
• Meta titles and descriptions generated for all pages after bulk-loading blog posts

The final phases were about building out the rest of the site and using Claude to speed things up wherever I could. Quick wins added up fast. It also got me thinking: when should we be building meta titles and descriptions into the workflow? Should it be treated the way we handle LiveInOkla blog posts — where everything comes to me pre-written and I’m just copying, pasting, and optimizing?

Chatbot Trance

One of the things I’ve noticed about working with AI is that if you really get locked in, it becomes a kind of trance — a state where you’re not quite aware of the decisions you’re making. The way I tend to use ChatGPT or Claude is mostly as a sounding board, a collaboration partner for a back-and-forth game of idea tennis. It helps me think through something. Over time it starts to feel like a brainstorming session with another version of myself — one that has the same starting knowledge, but whose next move I can’t quite predict. It’s a strange thing to describe, especially when you’re an hour or two deep in a conversation.

I’ve walked away from multi-hour Kiowa email sessions with copy for a dozen emails written — and I barely remember any of it. That’s the thing about working this way: we’re not used to thinking about how we’re thinking, or noting each decision we made along the way and why. I couldn’t possibly reconstruct why I made a particular writing choice on a Tuesday morning three months ago.

Walking away and coming back can be eye-opening. You return to the output you and your chatbot thought through together, and it often feels like someone else’s work:

• “I would have made a different decision.”
• “This sounded way better at the time.”
• “What was I thinking?”
• “This reads better flat than it does in a layout.”

It’s a weird phenomenon. Be aware of it — it feels pretty close to autopilot.

How the Professionals Do It

During work on this website, I watched a webinar by folks from an agency called Superside. It was called “12,000 AI Projects Later: The Creative Lessons You Can’t Google.” Over about thirty minutes, two people from Superside walked through how they loaded an LLM — Claude, in this case — with an extensive base of brand context: voice, mission and vision, content pillars, art direction styles, color usage. All of it compiled into source-of-truth documents that could be reloaded into any future project or chat.

The benefits they outlined were pretty clear:

4. Everything about the brand is already in the project — no re-priming at the start of each new conversation
5. Everyone who touches the project starts from the same foundation, which creates automatic alignment
6. Teams can work together with theoretically minimal conflict about how the brand should sound or look

Sound familiar? Superside was doing a more buttoned-up, standardized, and process-driven version of exactly what I’d stumbled into with Dragonfly. They had a name for it and a system around it. I had an intuition and a keyword bible CSV.

You learn something every project, and you adapt. Over the course of this one, I stumbled into the idea of building context — and midway through, I realized other people were doing the same thing, just with a lot more intention behind it.

Biggest takeaway: Context is everything. The more you have prepackaged in machine-readable formats like Markdown or CSV, the faster and easier it is to build up that brand brain — and the more consistent everything that comes out of it will be.

Note: As I’m writing this recap, it just dawned on me that everything up to this point was context-building in nature. Discovery is about learning context. Keyword strategy creates context. Each of these phases produced a canonical source of truth document that could be loaded into the Claude project to build up additional shared knowledge. I’ve been constructing the plane while flying it — which is a pretty wild thing to notice in hindsight.