During the first half of this year, I’ve worked with several organizations to roll out AI assistants and upskill their workforce. A few of the leaders have taken masterclasses and build-a-thons hosted by creators (often with much larger followings than mine). Yet, all of them shared how these sessions catered to founders, operators, and solopreneurs, but not much to real businesses and the leaders working within them. I’m not too surprised, and you probably aren’t either.
Most of the AI tutorials in our social media feeds include some form of email triage and automating LinkedIn posts. Few address the governance and processes that business users need. I quickly realized how much governance and scale are the red thread in projects like these, how little of it is shared online, and how much it reminds me of IT projects earlier in my career. So, this is my attempt to share four gaps I regularly see and how to close them to set your team up for success beyond the shiny demos. Getting Claude business-ready goes beyond quick tutorials, and these challenges only surface in real deployments.
(This article references Claude as the AI assistant. While it is the one I currently use, the concepts are equally applicable to other AI assistants.)
Gap 1: Addressing Version Drift
Tutorials walk you through the steps to teach Claude who you are and how you write. Some walk you through the process of creating Skills, but more often than not, it’s a fire-and-forget approach. The challenge is that the information you want Claude to reference for everyone evolves more quickly than you imagine at first. Take the example of a tone and style guide or a reference of your company’s products and services. As new products are added, others are renamed, or some are even being retired, that golden master file becomes outdated.
Org admins can make Skills centrally available to org members, or push out custom instructions to everyone on Claude’s Team plan (and above). But keeping information current often requires manually informing team members of updates and asking them to reinstall the required Skills. If team members don’t perform the update, version drift is pre-programmed, and Claude refers to incorrect terms, product names, or pricing information.
Solution: Separate instructions from data in a Skill
Not every Skill needs to have its entire reference data encapsulated. This approach allows you to maintain the information that changes more frequently compared to the instructions of how to use it, which change less frequently.
Instead, define the SKILL.md file to look up a central golden master, e.g., “brand-guide.md”, on Microsoft 365 OneDrive/ SharePoint (or any other centrally managed file storage). If your brand guide or product information changes, you only need to update that file while you users’ Claude instance always pulls the latest version.
[On OneDrive, SharePoint, or similar cloud storage]
assets/ # static information
└── fonts/ # corporate font files
└── logos/ # organization's logo files
kernel/
└── brand-guide.md/ # when to use which fonts, colors, and designs
[In Claude]
skill-name/ # actual skill
└── SKILL.md/ # instructions, referencing assets and kernel files
New LinkedIn Learning Course: Lovable 101: From Prompt to Product
Building a business app used to require a development team, a budget, and months of work. Today, AI-assisted tools like Lovable make it possible for any business professional to go from idea to published app in under an hour with no coding experience required.
In this course, instructor Andreas Welsch walks you through the complete process of building a Wins Journal app, a personal tool for logging professional achievements and generating an AI-powered year-end summary.
Along the way, learn how to validate an app idea, write effective prompts, add user authentication, store personal data securely, and publish your finished app. Walk away with a working, published app and the confidence to build your next one.
(Click the image to start the course on LinkedIn Learning…)
Gap 2: Ensuring File Integrity
If a centrally stored file is manipulated or tampered with, every user whose Claude Skill performs remote checks or file use will receive unwanted instructions that their instance processes. Balancing security risks and information governance is the real risk here. Addressing this situation would be a small deterministic piece of code. But when natural language is the programming language and the product lives across multiple surfaces, devices, and modalities, efficiency and security can decrease.
Solution:Establish strong information governance processes
Any centrally maintained files like your company’s brand guide, product information, or pricing details, must be the single source of truth if you expect every Claude user in your business to generate results grounded in your business data. Depending on the size of your company or team, maintaining this information could be an agenda point in a weekly team meeting.
Limit the number of team members who have write access to the central folder. Everyone needs read access, but only a few users can write. Ideally, create a user account just for the maintenance tasks to avoid any spillover effects if an admin account gets compromised or if their own Claude instance changes files by accident.
Thanks for reading The AI MEMO! This post is public so feel free to share it.
Claude guides users through the process of creating a new skill by using the /skill-creator skill. As part of this process, it proposes a name for the new skill. However, as your users’ skill repository grows, that naming frequently lacks consistency. Eventually, users spend time trying to determine which skill to use when. When sharing skills with other team members, the lack of a naming convention adds additional confusion where to find what, and when to use what.
Built and followed consequently, you can distinguish between skills that a user triggers, skills that other skills trigger, and the tier/ type of the skill. A naming guide helps you reduce the following:
Mis-trigger rate — wrong skill fires for a given intent. Target: below 5% of invocations.
Duplicate-skill rate — two skills doing one job because the author could not find the first. Target: zero net-new duplicates.
Time to first correct invocation — a new collaborator finding the right skill unaided. Target: under 60 seconds using the name alone.
Blast-radius legibility — share of skills whose write-scope is inferable from the name. Target: 100%, no exceptions.
Solution: Build your naming guide and taxonomy with Claude
As a best practice, define a naming guideline before rolling out any Skills to your team, so the naming is consistent from day one. You can even ask Claude to suggest different options for how to best set it up. Below is an example I use in my work. Claude’s actual naming guidelines are a lot more elaborate. You can even turn your skill-naming into a Skill of its own and apply it every time you create a new skill.
[tier-]<object>[-qualifier]-<verb>
Read right to left:
1) What it emits
2) What it acts on
3) Who is allowed to call it
Example:
lib-transcript-fetch co-onesheet-build newsletter-draft
└┬┘ └────┬────┘ └─┬─┘ └┬┘ └──┬───┘ └─┬─┘ └───┬────┘ └─┬─┘
tier object verb tier object verb object verb
The prefix answers who calls this and under what constraint.
There are exactly three tiers and only two prefixes; the default tier is unmarked.
| Prefix | Tier | Who invokes it | Constraint |
|---|---|---|---|
| *(none)* | **Top-level** | A person, by name or by intent | Must have a trigger-phrase block in its description |
| `lib-` | **Component** | Only another skill | Neutral: no voice, no brand, no opinion about content. Must be **reusable by** at least two callers |
| `co-` | **Brand-locked** | A person or an orchestrator | Output carries the COMPANY visual identity. **Must** load fonts, logos, palette, and voice guidelines from the `CENTRAL_FOLDER` folder before rendering |
Gap 4: Reducing Changelog Bloat
As you edit and evolve the Skills you use, Claude maintains a changelog at the bottom of a markdown file, indicating the key changes as well as when and why they were implemented. Every change adds an additional paragraph that grows and bloats your SKILL.md file. Every time the skill is loaded, unnecessary information is loaded into Claude’s context window and consums tokens. As a result, users will use up their allotted quota more quickly. Examples of this changelog look something like this (abbreviated to save space):
<!-- Cascade note (2026-04-03, batch 3 of the dependency-ordered rename programme): [...] -->
<!-- Rename provenance (2026-07-29, batch 4): podcast-ops -> podcast-run. Rationale under skill-naming.md v1.1: [...] -->
<!-- Cascade note (2026-08-25): implemented the Per-Step Production State spec [...] -->
Solution: Maintain a copy and cleanse the SKILL.md file in use
To separate the information Claude needs to process this skill from the history of how you have arrived at the latest version, separate the changelog from the Skill itself. Create a copy of the skill that has the changelog and ask Claude to create a clean version without the historic information. That approach reduces the number characters and tokens that Claude needs to process every time that particular skill is used. At the same time, you keep that history elsewhere for traceability and troubleshooting.
Summary
Influencers are littering your social media feed, creating fear of missing out (FOMO) and presenting easy-to-build demos. While the information can inspired new ideas and novel ways of using AI in your business, these foundational tutorials often skip the complexity of a real business that has several information sources, serveral versions of any given document, and several information creators.
Using AI well in a business is no longer just about acquiring prompting skills. IT and business leaders alike need to ensure information governance is in place, the integrity of information is maintained, and information is shared optimally. The approaches in this article are based on real-world deployments of Claude. They aim to close the gap in available information to scale Claude across teams and businesses.
Equip your team with the knowledge and skills to leverage Agentic AI effectively. Book a consultation or workshop to accelerate your company’s AI adoption.
Become an AI Leader
Join my bi-weekly live stream and podcast for leaders and hands-on practitioners. Each episode features a different guest who shares their AI journey and actionable insights. Learn from your peers how you can lead artificial intelligence, generative AI, agentic AI, and automation in business with confidence.
October 13 - Michael Nevski (Consumer Market Intelligence Expert) will discuss the prerequisites for AI agents to complete financial transactions on their user’s behalf. [More details to follow on my LinkedIn profile…]
Watch the latest episodes or listen to the “What’s the BUZZ?—AI in Business” podcast
