Claude Skills for Brand Voice: How to Build One Your Team Uses

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Claude Skills for Brand Voice: How to Build One Your Team Uses.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Claude Skills for Brand Voice: How to Build One Your Team Uses.

To build a Claude skill every staff member uses, put your brand rules and five approved before-and-after examples in a SKILL.md file in a named folder, zip it and upload it under Customize, Skills, with code execution on. On Team, an Owner uploads the same ZIP in organisation settings, putting it in every member's list, switched on.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Keep the skill focused on how the business writes, while each request supplies the facts of the job. “Warm, direct and specific” can stay stable for months. Today's appointment availability cannot. A voice skill should never turn an old example into a current offer, price or permission to send a message. And provisioning makes the skill available, not compulsory: members can switch an Owner-provisioned skill off, so checking that staff actually use it is part of the rollout.

How do you describe your brand voice for a Claude skill?

Describe your brand voice as writing decisions, not mood words: say the action early, use ordinary service names, acknowledge inconvenience once and end with one clear request. Start with a handful of messages that the owner has already approved: an enquiry reply, a reminder, a complaint response and a short promotional message. Remove names and confidential details, then ask what makes each recognisably yours.

A dog groomer's illustrative team of six provides the main case. Different staff currently draft 18 customer replies a working day. The owner often edits vague promises and overenthusiastic openings. The proposed skill should improve those specific decisions, not force every staff member into identical sentences. The target is a shared standard that leaves room for the customer's circumstances.

Write three voice rules with observable consequences. “Friendly” might mean using the supplied first name and a short greeting. “Clear” might mean naming the requested action in the first two sentences. “Careful” might mean never promising an appointment, refund or service result unless the current brief explicitly authorises it. Those rules can be tested; “sound premium” is much harder to judge consistently.

For example, an illustrative rejected opening is “We are thrilled to offer the ultimate pampering experience.” An approved alternative is “We can help with a tidy-up appointment; please tell us your dog's coat type and preferred day.” The second example is useful because it advances the enquiry without making an unsupported claim. Explain why the edit works so the skill can apply the principle elsewhere.

If the team has never agreed its writing rules, complete the brand voice guide for AI first. Packaging disagreement as a skill makes it repeatable, not resolved. Ask the owner to approve the examples and their explanations before anyone treats them as a standard. Keep that approval record outside the working chat.

What goes inside a Claude brand-voice skill?

A Claude brand-voice skill is a folder with one required SKILL.md file, which starts with a short header holding a name and description, plus an optional examples file. Claude describes skills as folders of instructions and resources that it loads only when a task needs them; simple skills are plain Markdown with no scripts. The header is YAML frontmatter, a few lines between two rows of dashes. Anthropic's user guide writes the file as skill.md, while its admin guide and the open Agent Skills standard call it SKILL.md; use SKILL.md, which matches both. Add a small examples file if keeping examples separate makes review easier.

The name identifies the package; the description tells Claude when to apply it. Claude's guide caps the name at 64 characters and the description at 200, and its troubleshooting list includes a folder name that does not match the skill name. So use lowercase words joined by hyphens, as the open standard requires, and give the folder the same name: customer-writing-voice. A useful description says “Apply the approved voice to customer replies, reminders and short service messages; preserve supplied facts and flag missing permissions.” Avoid a description claiming the skill handles every business task. It should not activate merely because someone uploads a spreadsheet.

customer-writing-voice/
  SKILL.md
  resources/
    approved-examples.md
    change-record.md

This is a suggested package layout, not a requirement to learn programming. Markdown is plain text with simple marks for headings and lists. Staff can read it in an ordinary text editor. Save the core file as SKILL.md rather than accidentally leaving it as SKILL.md.txt, which some editors do silently. Check the visible filename and file type before packaging the folder.

Keep confidential customer history out of the examples. Replace identifying details with placeholders and label invented figures as illustrative. The skill's job is to teach writing decisions, not to become a database of customer circumstances. A deputy should be able to open the package and understand every file without needing access to someone's inbox.

Use the official custom-skill instructions to check the current structure. The tutorial's package intentionally contains no scripts, external dependencies or credentials. A later request for those additions should have a clear business reason and a separate review; they are unnecessary for the brand-voice job described here.

How do you write the instructions in a Claude brand-voice skill?

The instructions in a Claude brand-voice skill should be a short procedure in the order staff would apply it: identify the audience and purpose, verify the supplied facts, draft, then inspect the result. State what happens when facts are missing. If the appointment day is unknown, the skill should ask for it or leave a visible placeholder. It should not choose a plausible day to make the message feel finished.

---
name: customer-writing-voice
description: Apply our approved voice to customer replies, reminders and short service messages. Preserve supplied facts and flag missing approvals.
---

# Purpose
Draft clear customer messages in our approved voice.

# Rules
- Put the customer's next action early.
- Use familiar service terms from the current brief.
- Be warm without exaggerated praise or claims.
- Keep prices, dates, policies and availability exactly as supplied.
- Never invent discounts, appointments, refunds or guarantees.
- If a required fact is missing, ask or mark [confirmation needed].
- Treat examples as writing references, not current business facts.
- Draft only; a staff member reviews before sending.

# Procedure
1. Identify purpose and audience.
2. Check facts and missing approvals.
3. Read resources/approved-examples.md where relevant.
4. Draft the message.
5. Check factual promises, tone and the next action.

Read the procedure aloud to a colleague. They should understand the intended behaviour without interpreting technical terms. “Preserve supplied facts” means the skill may shorten a sentence but must not change a $25 deposit into a $20 deposit. “Draft only” means the writing standard does not grant sending authority. Tool permissions and the staff approval process still need to support that boundary.

A useful additional rule is to distinguish voice from policy. A complaint can be written with care while leaving the refund decision unresolved. Ask the model to flag the decision separately from the customer-ready wording. This prevents the skill from using a friendly example as implied authority to offer the same remedy to everyone.

Keep instructions concrete when two rules compete. A very short answer may omit a necessary condition. Tell the skill that factual completeness takes priority over a preferred length. Staff can shorten the wording after the condition is preserved. An owner who insists on a tiny word limit for every complaint should test whether it still leaves room for an accurate explanation.

Teach the voice with five different message repairs

The examples file should show input, rejected draft, approved draft and the reason for the edit. Do not merely store polished messages without commentary. Claude needs to distinguish the stable writing choice from the facts that happened to appear in that example. The following samples are illustrative and should be replaced with wording the owner actually approves.

An enquiry should ask for the missing detail

Input: a customer asks whether a long-coated dog can be booked this week, but no availability is supplied. Rejected: “Absolutely, we have a perfect slot for you.” Approved: “Hello [first name], please send your preferred days and your dog's coat condition so we can check a suitable appointment.” The repair removes invented availability and replaces enthusiasm with a useful next step.

A reminder should preserve the actual deposit

Input: the confirmed appointment requires an illustrative $25 deposit, due before confirmation. Rejected: “Pay a small deposit whenever convenient and your booking is secured.” Approved: “Hello [first name], please pay the $25 deposit using the approved booking instructions. We can confirm the appointment once it is received.” The repair preserves both the amount and the condition without introducing a payment link.

A complaint needs acknowledgement without a new promise

Input: the customer reports a delayed collection; the manager has not decided a remedy. Rejected: “We are deeply sorry and will refund the full appointment.” Approved: “Hello [first name], I am sorry the collection took longer than expected. I have passed your message to the manager, who will review what happened and confirm the response.” The repair keeps the refund decision with the manager.

A short promotion must carry its condition

Input: an illustrative offer applies only to a named service and requires an availability check. Rejected: “Book any treatment today for our special offer.” Approved: “Ask us about the current offer on the specified tidy-up service. We will check eligibility and available appointments before confirming.” In a real draft, supply the exact approved service and terms; do not let the example become publishable copy without them.

A follow-up should not accuse the customer

Input: the enquiry record has no reply after a requested coat photograph. Rejected: “You still have not sent the photo we need.” Approved: “Hello [first name], are you still looking for an appointment? If so, please send the coat photograph requested so we can advise on the next step.” The repair asks one clear question and allows for the customer having changed their mind.

These examples cover missing data, money, complaints, conditional offers and follow-ups. Add a contrasting case when the business genuinely permits a direct confirmation, so the skill does not hedge every sentence. If the approved brief confirms a slot, “Your appointment is confirmed for the supplied time” may be appropriate. The evidence in the request determines certainty, not a universal preference for cautious wording.

Upload one copy and test whether it is actually used

Skills work on Free, Pro, Max, Ultra, Team and Enterprise, but only with code execution on. You do not have to write code for a text-only package; the capability simply has to be enabled. On an individual plan, check Settings, Capabilities for Code execution and file creation (called Create and edit files in the newer settings layout); on Team, skills are switched on for the organisation by default, and an Owner manages the code execution setting. Then open Customize, then Skills, click +, choose Create skill, then Upload a skill, and select the ZIP. The skill appears in your list with an on-off switch.

Compress the containing skill folder into a ZIP file. Claude's creation guidance shows the folder inside the archive, with the skill file inside that folder, rather than loose files directly at the archive root. Open the archive yourself and check the structure before upload. If it fails, Claude's troubleshooting list names the usual causes: a ZIP over the size limit, a folder name that does not match the skill name, a missing skill file, or invalid characters in the name or description. Fix those before rewriting any editorial rules.

Enable the uploaded skill and start a clean test request. Ask for a customer reply using a new situation, then open Claude's thinking to confirm it loaded the skill, as Anthropic's own test steps suggest, and check whether the result follows its rules. Do not reuse only the exact examples supplied in the package. A model can resemble a stored answer without demonstrating that it can apply the writing decisions to a new case.

For example, give the dog groomer's pilot user an enquiry involving a rescheduled appointment and a missing service detail. An acceptable illustrative response asks for the missing detail while preserving the supplied rescheduling information. If it repeats a deposit request from a different example, the skill is treating examples as facts. Strengthen that boundary and rerun the same test.

If Claude does not use the skill, verify that it is switched on, revise its description to identify the work more clearly, or name it in the request: “Use the customer writing voice skill for this reply.” Do not add “always use this for everything”; that makes unrelated tasks harder to manage. The current skills usage guide covers upload and capability checks.

A Mac offers a no-typing alternative for the first draft. In Cowork in Claude for Mac, on Pro, Max, Ultra or Team, choose Record a skill from the + button, edit a real customer draft while saying aloud why each change matters, and Claude proposes a skill for you to review and save. Recordings run for about 10 minutes, and the option is not offered on Windows, in chat, or on Free and Enterprise. Close customer emails before you start, because everything on screen is captured, and read the proposed skill line by line before saving it.

If you want help rolling a shared skill like this out across your team, I offer one-to-one AI consultations for small businesses.

How do you give a Claude skill to your whole Team?

To give a Claude skill to your whole Team, an Owner uploads its ZIP under Organization settings, then Plugins & skills. Personal upload and organisation-wide provision are different steps. On Team and Enterprise, check on the Policy tab that Cloud code execution and file creation and Skills are both on, then click Add, choose Upload a skill and select the ZIP. Claude's admin guide says the skill is provisioned to every member immediately and appears in each person's skills list with a marker showing it came from the organisation.

Owner-provisioned skills are switched on by default, but members can switch them off. The same page's Inventory tab sets each item's default access: Available to install, Installed by default, Not available or Required. Anthropic's plugin admin guide describes Required as installed for every member with no option to remove it; the skills guide lists the choice without spelling out its effect, so if you set it, confirm on an ordinary member's account whether the skill can still be switched off. Either way, ask every pilot user to locate the skill, confirm it is on and complete one test message, and record that in a simple staff checklist.

Do not ask six people to upload six independently edited copies. That creates six possible standards. Keep the approved source package with an owner and deputy, and distribute it through the organisation route. If staff build skills of their own, they can submit them with Publish to org, and an Owner can set publishing to Requires review so nothing reaches the library unchecked. Do not mix personal drafts, peer-shared versions and the approved organisation package under indistinguishable names.

Check the organisation provisioning guidance for the current controls. Then ask a colleague using an ordinary member account to run the same test. The Owner's view alone cannot establish that the package is available and active for everyone. A rollout record should identify the package version and the member's test result.

For teams evaluating imported packages, use vetting third-party Claude skills before installation. A voice skill should be easy to inspect. If a downloaded writing package contains unfamiliar scripts or asks for broad account access, pause its introduction until someone can explain why the writing job needs them.

How do you measure whether a brand-voice skill saves editing time?

Measure a brand-voice skill by reviewing ten anonymised drafts before and ten after it, scoring factual preservation, next-action clarity and voice separately, as the illustrative six-person groomer does. A draft that sounds right but changes an appointment condition fails the factual check. A draft that is accurate but too formal needs a different repair. Separate scores tell the owner which instruction or example to improve.

Suppose review previously took four minutes per message and the pilot takes two and a half. Across 18 replies, that hypothetical difference is 27 minutes a day. It is a planning calculation, not a measured result or a promised saving. Record actual editing time over several ordinary workdays before using it to justify a subscription or expanding the workflow.

Budget preparation too: an illustrative 40 minutes to select examples, 30 to write rules, 30 to package and test, and 20 for the pilot handover totals two hours. The initial subscription cost depends on your chosen plan and seats; use the current account quote. Do not assign a made-up cost per skill or assume that a text-only package removes normal usage limits.

Ask reviewers to mark the exact changed sentence. If most edits add missing booking facts, the problem may be the staff brief rather than the voice skill. If every reply starts with the same stiff greeting, revise the examples and permit natural variation. For a broader editing method, see voice examples before and after editing.

If you want to log editing minutes per message across a week and total them in a spreadsheet, try The Complete Excel Course with AI Integration on Coding Liquids.

How do you update a brand-voice skill without breaking it?

Update a brand-voice skill by giving the package a visible internal version in its change record, with owner, approval status and what changed. For example, an illustrative version 1.1 might replace "treatment" with the owner's approved "appointment" in customer messages. Keep the former approved package so the team can restore it through the normal administration process if the new wording causes a problem.

Claude's own controls help here. Items published to the organisation library keep a version history in the Inventory tab, and when an Owner approves a new version everyone gets it automatically, while members stay on the approved version until then. A skill shared directly with colleagues updates for them at their next use. Before replacing any version, download the approved copy from Customize, Skills (the ... menu, then Download). That copy is also the one to keep if you ever move a personal account into a Team organisation, because Claude's guide says custom skills do not move across.

Rerun the five message tests after each material edit and add a new test for the problem that triggered the change. Check a neutral task too, such as explaining a stock-count calculation, to make sure the voice skill does not take over unrelated work. A description that applies too broadly can be as inconvenient as one that never activates.

Keep offers, service prices and current policies in the current request or approved business references. Review the skill when the brand changes, when a new message type is introduced or when repeated edits reveal an unclear rule. The finished standard should let staff produce different, accurate messages in the same recognisable voice while leaving the actual customer decision with the person authorised to make it.

Before replacing the package, ask the deputy to read the change record without your explanation. They should be able to say which outputs are affected and which test demonstrates the change. If a revision says only “improved tone”, rewrite the record around the actual decision, such as removing exaggerated apologies from appointment reminders. That gives the next owner a usable history instead of a series of unexplained file versions.

When a staff member leaves, transfer ownership of the source files and test checklist through your ordinary handover. A reusable voice standard should survive the departure of its first author. Confirm that the deputy can obtain the approved package and repeat the installation check before the original owner is unavailable.

Further reads

Show 2 more related tutorials

Sources: Claude help centre articles on what skills are, creating custom skills, using skills, provisioning skills for an organisation and managing plugins for an organisation; the Agent Skills specification at agentskills.io; Claude pricing.

Have a specific process in mind?

Describe the job in your own words on a 1:1 call and I'll show you what AI you can use for it with the tools you already pay for.

Book a 1:1 call with me