Give it your brand voice

The problem: a generic tone

The /post skill works, but its output sounds generic. Every platform gets the same tone: the Twitter posts read like LinkedIn, the LinkedIn feels too casual, and all of it could have been written by any brand in any industry. For Lea, that's a deal-breaker: her brand has a voice — warm, committed, never preachy — and her customers recognize it. Content that doesn't sound like "Lea" is worse than useless: it dilutes the brand.

This problem isn't a bug, it's a property of language models. An LLM is trained on the entirety of available text: without instructions to the contrary, it produces the statistical average of what it has seen — a one-size-fits-all tone, consensual phrasing, predictable structure. An AI's default voice is nobody's voice.

Why AI writes "like AI"

You've surely spotted the tics already: cascading em dashes, three-beat formulas ("not just X, but Y and even Z"), inflated words like "revolutionary," "dive into," "unlock the potential," conclusions that restate what was just said. These markers come from the model optimizing for "well-formed" text in the schoolbook sense — balanced, exhaustive, polite — whereas a good social post is the exact opposite: sharp, incomplete, with a real personality.

The practical consequence: you don't fix a generic tone by asking it to "write more like a human" (vague instruction, vague result). You fix it with two concrete levers: explicit negative rules (the list of what's banned) and real samples (the model is excellent at imitating a style you show it — far better than at guessing a style you describe).

Each platform has its own grammar

Beyond the brand voice, every network has its codes — and mixing them up is the most visible sign of carelessly automated content:

Your skill therefore needs to encode two layers: the brand voice (constant everywhere) and each platform's grammar (variable). That's exactly what the next prompt does.

The "humanize" prompt

add my brand voice samples and my style to the /post skill:

# STYLE (all platforms):
- clear, simple language, short and punchy sentences
- active voice, not passive
- practical, actionable insights
- direct address: "you" and "your"
- AVOID em dashes, metaphors, cliches, generalizations
- AVOID catch-all words (revolutionary, disruptive, powerful, dive, unlock...)

# SAMPLES PER PLATFORM:
Twitter: <PASTE 3 EXAMPLES>
LinkedIn: <PASTE 3 EXAMPLES>
Instagram: <PASTE 3 EXAMPLES>

Paste real posts you've written — your best ones, the ones that performed well or that you're proud of. The more examples you give, the better Claude matches your tone: three per platform is a good start, five or six is even better. The style rules at the top eliminate the markers that betray AI text; the samples provide the positive signal that rules alone can't give.

An organizational tip: rather than cramming everything into the SKILL.md, ask Claude to create a dedicated brand-voice.md file that the skill references. The advantages: you can update the voice without touching the skill, the same file will serve other skills (the /plan-week of chapter 7), and a reference file loaded on demand keeps the skill lightweight.

No samples yet?

If you have no samples — brand-new brand, career change, new network — ask Claude to go find some: "go find 10 writing examples from [an author you admire] on LinkedIn." It will use them as a voice reference. Pick someone whose tone genuinely matches what you're aiming for, not just someone famous: you inherit their tics as much as their qualities.

Think of it as temporary scaffolding. As soon as you publish and some posts perform, replace the borrowed examples with your real posts. Over the weeks, your sample library becomes your most precious asset: the operational definition of your brand.

Iterate methodically, not at random

The first version won't be perfect — that's normal and that's the plan. What matters is iterating with method. Generate the same topic on two platforms, read the outputs aloud (the most reliable test for detecting "AI tone"), and identify precisely what's off: a word, a turn of phrase, a length. Then fix the rule, not the post: "add the word elevate to the banned words list," "Twitter posts exceed 200 characters, tighten the constraint."

flowchart LR
  G["Generate a test post"] --> L["Read aloud"]
  L --> D{"Something off?"}
  D -->|"Yes: identify the exact tic"| R["Fix the rule in the skill"]
  R --> G
  D -->|"No"| V["Voice validated"]

The big lesson: a skill evolves

Step back and look at what we just did: we didn't replace the skill — we augmented it with context. This is essential: your skills should never be frozen. Like a human competence, you improve them continuously: what works, the edge cases you encounter, your preferences as they sharpen. Every improvement benefits all future uses — it's a compound-interest investment.

Concretely, get into the habit of saying "update the skill so that…" every time an output disappoints you. Fixing the output repairs once; fixing the skill repairs forever. That habit — improving the tool rather than the deliverable — is the through line of this whole course, and you'll meet it again in chapter 7 as the central pattern.