How we built brand-aware summaries

When a generic AI reads a whiteboard, you get a generic summary. It uses words like "deliverables" and "stakeholders" and "action items" in a flat, universal voice. That's fine for a demo. It's not fine for a consultant who spent months building a client relationship, or a startup with a specific brand POV, or a product team that has strong opinions about how they talk about their roadmap.

Brand-aware summaries were the feature that turned BoardSnap from "a whiteboard scanner with AI" into something that felt like it actually belonged to you.

Here's how it works.

### The problem with generic AI output

Let me be specific about what "generic" means in practice.

Imagine a fintech startup with a brand voice that's skeptical of incumbent banks, direct with customers, and uses plain language intentionally. Their whiteboard from a product meeting might include phrases like "kill the friction" and "real people, real money." A generic AI summary of that board would output something like: "The team identified opportunities to improve the customer onboarding experience and reduce friction in the payment flow."

That's not wrong. But it sounds like it came from McKinsey, not from them. Nobody on that team would write "opportunities to improve" — they'd write "three things we're fixing." The summary is accurate but alienating.

Brand voice isn't decoration. It's how a team recognizes its own thinking when it's reflected back at them.

### How we crawl the brand

When you create a Project in BoardSnap and add your website URL, we fetch and parse the site. Specifically:

1. We pull the homepage HTML, strip layout and navigation, and extract the visible prose content.
2. We pull up to two additional pages — typically the About page if we can find it, and one product or feature page. This gives us vocabulary, not just positioning.
3. We run a distillation pass to extract: tone descriptors (formal/casual, technical/plain), vocabulary patterns (recurring nouns, preferred verbs), messaging pillars (the 3–5 claims the company makes about itself), and any explicit persona signals ("we are" / "we believe" / "our approach is").
4. The distilled brand context — typically 200–400 tokens — gets stored per Project.

This happens once, on Project creation. Subsequent snaps in the same Project use the cached context. The crawl adds maybe 3–5 seconds the first time you set up a Project. After that, zero cost.

### How it gets injected

Every BoardSnap summary prompt includes the brand context as a system-level instruction:

`` You are summarizing a whiteboard for [Company Name]. Brand voice: [distilled descriptors] Key vocabulary: [recurring terms] Messaging pillars: [key claims] Write the summary as someone from this company would write it. ``

The AI isn't copying the website. It's adopting the voice, the vocabulary, and the attitude. A boardroom whiteboard summary for the fintech startup above would read more like: "Three friction points to kill before the Q3 push: onboarding step 4, payment confirmation, and dispute flow. Here's who owns what."

Same facts. Completely different feel.

### What it gets right, and where it struggles

Brand-aware summaries work best for companies with a clear and consistent voice — startups with a strong POV, consultancies with a distinctive methodology, agencies with a recognizable style. The more distinct the brand, the better the results.

They work less well for:

• Companies with generic corporate copy. If your website reads like every other company in your space, there's not much to extract. The summary will still be better than fully generic, but the delta is smaller.
• Technical teams. If your brand voice is really just "we use precise technical language," that translates — but it's less dramatic than a brand with a distinct personality.
• Multilingual brands. We currently only extract English-language copy. If your brand speaks primarily in another language, the context extraction isn't reliable yet.

### The pinned context layer on top

Brand voice is one layer. Pinned context is a second layer.

You can pin specific notes inside a Project — your team's preferred frameworks, your current OKRs, the terminology that matters right now. Pinned notes get injected into every chat and every summary alongside the brand context.

The combination is powerful: the AI writes in your voice and knows your current priorities. A summary from a Q3 planning board doesn't just sound like your company — it references the OKRs you've pinned, connects action items to the goals you've set, and uses the framework your team has committed to.

This is the gap between "AI that summarizes" and "AI that's actually in the room with you."

### One thing I'd do differently

The current brand crawl is static — it happens once on Project creation and caches. If you rebrand, or if your messaging evolves, you have to manually refresh the crawl.

A better version would periodically re-crawl (maybe monthly) and diff the brand context against the cached version. Significant drift would prompt a refresh. Stable brands would never feel it. I have this on the roadmap but didn't ship it in v1.

For now: if your brand voice has changed meaningfully since you set up your Project, delete and recreate the Project to get a fresh crawl. It takes 30 seconds.