My Brand Skill Now Corrects Itself. Here's What That Actually Means.
My Brand Skill Now Corrects Itself. Here's What That Actually Means.
By Cynthia Schomp · Fractional CTO
Every brand skill starts drifting the moment you stop watching it.
I've been running Claude skills for months. Skills that load my brand system, my voice rules, my banned words, and write copy that sounds like me. They work. Until they don't. Until three sessions later the output starts feeling a little generic, a little AI-generated-for-a-marketing-page, and I'm back to correcting the same patterns I already corrected two weeks ago.
That's a pesky memory problem.
Three Files That Change Everything
I added an adaptive learning layer to my brand skill. Three reference files that Claude reads before touching any copy.
voice-rules.md: the core mechanics of how I write. First line equals the point. No windup. Short sentences. Specific claims over impressive-sounding ones. Senior practitioner briefing a peer, not a marketer writing at a prospect.
banned.md: every word and pattern I've explicitly flagged. Not just the obvious ones. The subtle ones too. The structural clichés. Anything hollow that sounds like it came off a SaaS template.
approved.md: lines I kept without changes. Phrases I said yes to. These train the system on what right actually looks like, not just what wrong looks like.
The skill reads all three before writing. Every time.
The Loop That Makes It Actually Adaptive
The files aren't static. When I flag something, the skill updates banned.md immediately and saves a feedback memory that persists across sessions. When I approve copy without changes, approved.md gets a new entry.
The system teaches itself on my actual preferences. Not my stated preferences. My revealed ones.
That's the difference between a style guide nobody reads and infrastructure that runs. A style guide is documentation. This is a feedback loop with memory.
Why This Matters Beyond My Own Brand
Every agency running AI-assisted content production has this problem. The first output is decent. The second is fine. By the sixth session the model has drifted back to whatever it thinks a professional marketing voice sounds like.
The fix isn't better prompts. It's captured feedback. Running lists of what failed, what landed, what the client actually kept when they stopped editing. That's the training data you're not collecting.
I'm collecting it now. In a format Claude can read every time it opens the skill.
How to Build This for Your Agency
- Start a voice-rules.md file. Write the three rules that matter most for your brand. Not twenty. Three.
- Start a banned.md file. Add words the moment you catch them. Don't batch it.
- Start an approved.md file. Copy the lines your client kept without editing. Ground truth.
- Wire all three into your brand skill as required reads before any copy output.
- Make the skill update the files when feedback happens. Not planned. Immediately.
The skill that learns from you gets better. The one that doesn't gets fired after three months.
What I'm Still Figuring Out
The approved examples file is the hardest to maintain. Banning words is easy. Recognizing a line that's exactly right takes a half-second more attention. I'm building the habit of calling it out when it happens, not assuming I'll remember it later.
I won't.
If you missed any emails in this series, the full archive is waiting: hellocyn.com/archive -- every issue, in order, yours to read whenever.
-- Cynthia Fractional CTO · Schomp.ai · The CTO's Desk
Reply to this if you're building something similar. I read every reply.
Get the next one.
One email per day. Real deployments. No theory.