Tallo Brand System

Tallo’s brand has to land with a 14-year-old figuring out what careers exist and with the district administrator who brought Tallo to their school. Stakeholders were drafting that copy with AI that had never seen our brand, and we spent our time fixing it. I rebuilt the brand as structured data, generated context files for each role (the instructions an AI reads before it writes), and built Claude workflows, a brand linter, and an eval on top, so the first draft starts on brand. About 15 people across the company use it, and it was built to keep running after I left.

See also: Cal’s Career Quiz, including the quiz system and Cal Kit.

The right word changes by room. “Students” works for educators and is wrong in consumer copy. “Emerging talent” belongs in an employer pitch and nowhere else. Some of the audience are minors, so every photo and clip carries consent rules too.

All of that lived in two long PDFs that nobody read mid-task, and the AI had never seen them. Meanwhile, stakeholders had started drafting copy with AI and handing it to us as a brief: roughly what they wanted, for us to shape from there. Those drafts came in with invented stats, retired terms, and the wrong voice, and shaping them often meant starting over. Brand review ran through two people, me and one writer, and we kept rebuilding the same decks from scratch. If AI was going to write the first draft anyway, the first draft should start on brand.

I rewrote the brand as structured data: colors, type, voice, the mascot, and the only approved list of stats. From that one source I generated a shared core file and a context file for each role, from social to exec, then built Claude skills and process templates on top. Claude, our n8n automations, the team’s Canva kit, Figma, Paper, and our developers all draw from it.

It built on groundwork from the year before: a digital asset library where every file is tagged by audience, channel, and consent status, including whether a student release form is on file. The brand system did the same for words, so audience and consent rules ride along with the content.

brand-system/├── brand/│   │   # source of truth, as data│   ├── colors.json│   ├── typography.json│   ├── personality.json│   ├── mascot.json│   ├── value-points.json│   │   # the only place a stat may come from│   └── social/│       # social palette + contrast decisions├── claude-templates/│   ├── BRAND.md│   │   # shared core: voice, terms, banned phrases│   └── *-context.md│       # one per role: social, blog, copy, design, dev, events, exec├── process-templates/│   # challenges, video production, presentations├── setup.sh│   # 15-minute onboarding├── QUICKSTART.md│   # for people who've never opened a terminal├── QUICKSTART-COWORK.md│   # the same setup in Claude Cowork, no terminal├── CONTRIBUTE.md│   # how to propose changes, what can ship└── CHANGELOG.md    # what changed, where, why, evidence

I gave a flat palette a hierarchy. I inherited 38 swatches at equal weight and a pairing chart with two failing pairs in its approved column. I sorted them into three tiers, gave each color family a job, built light and dark ramps, and tested every pairing I documented.

Inherited

38 swatches, all the same weight

Rebuilt

Three tiers, and every family has a job

Primary · the foundation, most touchpoints

• Nightshade#1E0F41
• Purple#7960DC
• Lilac#E2D6FF
• Green#BDF184

Secondary · info and interactions

• Navy, #0C2A86
• Blavy, #0D45B5
• Blue, #0D6EFD
• Sky, #ADE2FC

Education, informational

Tertiary · stays in its family

• Evergreen, #07362E
• Panache, #F8FED3

Finance, opportunity

• Orange, #FF6D5B
• Rust, #5D0E04
• Plum, #5C194D
• Salmon, #FFD6CC

Highlights, deadlines

Every stat has a paper trail. value-points.json is the only approved source for a number. Pulled stats move to a retired block with what replaced them. When an employer count turned out to be months out of date, the fix happened once and every tool picked it up.

"retired": [
  {
    "value": "500+ employers and organizations",
    "replaced_by": "900+ employer partners",
    "retired": "2026-08-05"
  }
]

Consent decides what ships. Before video transcripts went into the shared system, I replaced “is it already online?” with three tiers: always fine, case by case, never. Anything involving minors is never. A classroom recording and a one-on-one interview can look identical in a folder. The guest signed a release. The students in the room didn’t.

I build with Claude Code daily and run it like a small team. A stronger model plans and judges, a cheaper one carries out the plan, and the judge reports only what would block shipping, so I decide what to fix.

The brand system works the same way. The AI gets context scoped to the job, a short list of hard rules, and a check wherever one exists, like the deck skill’s contrast audit and page screenshots. What’s true, what’s approved, and what’s safe to publish about young people stay with people, and the system makes that explicit.

Most of the team doesn’t use a terminal, so getting started was the hard part. setup.sh asks your role and scaffolds a working setup in about 15 minutes. Claude Cowork users get the same setup with no terminal.

Claude-Context/
├── ABOUT ME/
├── PROJECTS/
├── tallo-brand-localsystem/
├── CLAUDE.md  ← Social media
└── global-instructions.md

Then the work started coming back. Two of these three pieces were made by teammates, for three different audiences, and all of them read as Tallo.

1. , enlarge

   Instagram post · for young people

   Made by our social media manager

2. , enlarge

   LinkedIn ad · for job seekers

   Made by our social media manager

3. , opens on YouTube in a new tab

   Educator video · for educators

   Started by our social media manager, finished by me

Most needed only small tweaks, a quick review where drafts used to mean starting over. The video took more work to finish for YouTube, but I didn’t start from scratch, and the social media manager’s cut already worked for a student hackathon showcase on a tight deadline.

Process templates carry work that used to live only in my head. Our community challenges follow a four-week arc: three episodes, then a closing reflection question. After running it twice, I wrote it up as a guide. A writer used it to plan a new month’s series on their own, and I built the challenge from their plan.

Watch the episodePlay: Ep. 1 | Would You Actually Want to Be Your Own Boss? ft. Real Founders

The oldest template is for video. I wrote our video production guide in December 2025, and our social media manager and interns have made most of our TikToks from it since. Our most-watched video reached 1.4 million views. A candid, funny moment is why it spread, but it went out in our brand’s caption style, so it still read like Tallo even though it came from the intern’s perspective. I made a few videos myself to learn the team’s process, then folded what I learned back into the guide and template files.

The same files power smaller tools. My alt text kit, a Claude project, writes descriptions that are accessible first and search-friendly second. It stays under 125 characters, spells out acronyms so a screen reader says “Workforce Innovation and Opportunity Act” instead of “W-I-O-A,” and always describes Cal as a compass, never as holding one.

Context files only guide the AI, so I built two ways to check them: a linter and an eval.

The linter reads its rules straight from the brand files: banned phrases, retired terms, any number not in value-points.json, em dashes, and reading level. Because it reads the source, it can’t drift from it, and it fails a draft the way a test fails a build. Its first run flagged “emerging talent” in an employer pitch, where it’s the right term, so it now takes an audience setting.

For the eval, I wrote 15 requests the team actually sends, several built to bait a known mistake, like “include a few impressive numbers.” Each ran three times without brand context and three times with it, isolated from my own setup.

Asked for a LinkedIn post about how big Tallo has grown, the baseline invented a different user count every run:

• “...a thriving community of over 5 million students and young professionals...”

• “📈 2M+ students now on Tallo...”

• “📈 3M+ users on the platform”

With the system, all three runs used the same approved figure:

What the numbers don’t say. The linter checks the same rules the context teaches, so this shows the AI follows the rules. Whether the copy is good still takes a human read. Fifteen requests is a directional signal. Reading level improved but missed our grade 5 target.

A brand guide only one person can edit expires when that person leaves. So I built the way in alongside the system. CONTRIBUTE.md says how to propose a change, what ships without review, and what needs a brand call. QUICKSTART.md gets a new hire running in 15 minutes without a terminal. And Christian co-maintains it from inside his own n8n and SEO workflows, so the system has an owner who didn’t inherit it as a chore.

It works. Christian noticed the role templates and our automation file each restated the brand voice in their own words, and the copies had started to contradict each other. One listed “empowering” as a core trait while another banned “empower.” His fix made BRAND.md the only place voice is written in prose. It’s the single-source rule I’d applied to stats, now applied to voice, and I didn’t have to be the one to spot it.

One thing I’d change: for months the system lived only in Bitbucket, where most of the people it was built for never looked. Dated release zips on the shared drive fixed access but split the system in two, because changes from drive users came back as messages someone had to turn into pull requests. A contribution path only counts if it reaches the people doing the work. I should have designed both from the start.

In testing, the system took made-up stats to zero. Then a flyer drafted with it, which skipped the linter, came back reading “Just 1 in 7 young people say a real conversation with a pro helps most.” The report behind it was on our approved list, but the survey let people pick up to three sources, and the top answer was family, friends, or mentors at about 1 in 3. We decided to pull the research from the repo to keep unapproved numbers out, so nothing checked how its numbers were framed.

Next time, every stat carries what was asked, where it came from, and who it’s for.

The second lesson is about where brand and product meet. Tallo’s marketing and product grew up as two systems. On social, we learned a purple button disappears in a purple frame, so green marks the one thing to act on. Lessons like that lived in marketing’s files, and product made its own calls in its own system.

In “Brand as Software (opens in a new tab),” Paul Jun of Ramp argues a brand should run like a product. The line I mentally highlighted is “Taste remains human. Repetition becomes software.” This system took on the repetition, and the flyer showed what still needs a person. What it becomes next belongs to the team: give product a seat at it, so both sides draw from one set of decisions. That’s when brand and product can play a stellar duet.