Check out our upcoming events | Learn More

AI agent skills: Beyond the platform

The new way to build mini applications on AI: what a skill is, when to use one instead of an agent, and how to write, evaluate, and govern them across every platform your teams use.

  • Navan Tirupathi
  • October 21, 2026
  • 6 Minute Read
Back to All Insights

KEY TAKEAWAYS

  • A skill encodes how your firm does a task, once, for everyone. A folder of plain-language instructions, no code, that makes AI output consistent instead of varying by whoever happened to prompt it.

  • Encode preference before capability. Capability gaps close on the vendors’ roadmaps; your judgment and standards are in nobody’s training data, so the skills that encode them are the durable asset.

  • One format, every platform, but govern it. The SKILL.md format is read by roughly fifty tools across competing vendors, so the library travels with you. Treat skills as versioned, owned, registered configuration, not a drawer of documents.

Most organizations have adopted AI. Far fewer can point to results that show up in the numbers. The gap is rarely the model. It is that AI use stays personal: every employee prompts their own way, quality varies with the person, and nothing anyone learns is kept.

Skills close that gap. A skill is a small folder of plain-language instructions that teaches an AI agent to do a specific task your way: your method, your standards, your output format. Written once, it works for everyone, every time. It behaves like a mini application, except no one wrote code, no one filed a ticket, and it took hours rather than months.

In under a year, the skill format Anthropic published became the shared convention across competing platforms: OpenAI, Microsoft, Google, Cursor, JetBrains, and roughly fifty tools in total read the same folder of markdown.1 Which means the skills your teams write are not a bet on one vendor. They are a durable asset that travels with you.

What follows covers what a skill contains, when to use a skill versus an agent, how the cross-platform landscape actually stands, and how to write, evaluate, and govern skills so the library stays an asset rather than a drawer of documents.

The AI platforms such as Claude, ChatGPT, and Microsoft Copilot will all keep changing. Your firm’s methods, standards, and judgment do not. Skills are a new way to encode them once and carry them anywhere.

A skill is a folder. Inside it sits a markdown file called SKILL.md that names the skill, describes when it applies, and then explains, in plain prose, how to do the task. The folder can also carry scripts for the deterministic steps, reference documents for standards and examples, and assets such as branded templates. That is the entire mechanism, and the simplicity is why it spread.

The design detail that makes skills viable at scale is progressive disclosure. The agent carries only each skill’s name and one-line description until a task matches; only then does it load the full instructions. An organization can hold hundreds of skills without slowing anything down.

Exhibit 1

Not all skills hold their value equally, and the distinction should drive where you invest.

Capability uplift. Gives the agent an ability it lacks, such as producing a polished document or running a data pipeline. Useful, but every capability gap you close by hand is one the vendors are also closing on their roadmaps.

Encoded preference. The agent can already do the task; the skill specifies how your firm does it. Review standards, brand rules, proposal structure, quality thresholds. These are durable, because your judgment is in nobody’s training data and never will be.

Encode preference before you build capability. It is the less glamorous investment and the one that compounds.

Exhibit 2

Leaders hear both words constantly and the distinction is simpler than the noise suggests. A skill is packaged know-how: it tells an agent how to do a task. An agent is a worker: it pursues a goal across many steps, often unattended. One is a recipe; the other is a cook.

The practical consequence: most enterprise value today sits in skills, because most valuable work is repeatable work with a known output. And skills are where the leverage multiplies in a collaborative workspace like Claude Cowork: the skill runs inside your session, with your files, your connected systems (email, SharePoint, Slack, CRM), and your permissions already in place. Nothing to integrate, nothing to deploy. The same skill that formats a document can pull the source data from the systems you are already signed into.

The skills you write today power the agents you deploy tomorrow. An agent equipped with governed skills inherits your standards; one without them improvises.

The strongest reason to treat skills as infrastructure rather than a productivity tip is convergence. The SKILL.md format is now read natively by the major AI platforms and a long tail of developer tools. Teams do not need to be on the same vendor to share the same library, and not every workflow needs to run in Claude.

Exhibit 3

Two honest qualifications belong next to that picture. First, the file travels but behavior needs light adaptation: platforms discover skills from different directories and differ in how they run bundled scripts, so budget a small per-platform validation step rather than assuming zero-touch portability. Second, support on the most widely deployed enterprise surfaces is still arriving. Microsoft 365 Copilot declarative agents read skills only in preview, Copilot Studio only for its newer agents, and Google is folding Gemini Gems into skills over the coming months.2 Treat those surfaces as arriving, not ready.

What this means for lock-in: adopting the format is low-risk. It is published under a permissive open license, and the asset you accumulate, a library of your firm’s methods in plain text, moves with you if platforms change.

A skills program lives or dies on a simple lifecycle. The pattern that works is the one platform teams already know: central standards, federated authoring, and a single internal registry.

Exhibit 4

Writing Well

Anthropic’s published authoring guidance reduces to a few rules that any team can apply:

Be concise. The agent is already smart. Only add what it cannot know: your context, your rules, your format. A tight one-page skill beats a ten-page manual.

Name and describe precisely. The description is how the agent decides when to use the skill among hundreds. Say what it does and when it applies.

Match freedom to fragility. Give the agent latitude where many approaches work; give exact steps and scripts where consistency is critical.

Keep it self-contained. Templates, standards, and scripts live inside the folder, so the skill works on any machine and any platform.

Evaluating Before Trusting

Test a skill the way you would test software you depend on: define what good looks like before authoring, run it against real cases including the awkward ones, and compare against a baseline without the skill. Do not accept enthusiasm as evidence. Baseline two workflows, measure completion time and rework, and scale on the measured result.

Governing Without Bureaucracy

A skill that shapes how real work gets done is configuration, not documentation. The control set is familiar and light:

  1. Version control, not shared drives. Every change is a reviewable diff with an author and a date.
  2. Named owner and review date. Skills without either get retired on a cadence. This is what prevents the library becoming a document graveyard.
  3. One internal registry. Skills enter through a reviewed intake path. Public-registry skills are reviewed before use, without exception, and the sanctioned path is made faster than the unsanctioned one.
  4. A curator. One accountable person owns library coherence: overlaps, drift, retirement. This role, more than any tool, determines whether the program scales.

Skills are the third level of a five-level progression, and the transition into them is where the payoff concentrates: the moment a personal prompt becomes a versioned, governed, shared asset.

Exhibit 5

Skills, as pioneered by Claude, are now arriving on every major platform, and each one packages them differently. As of October 2026:

PlatformSkill formatKey toolsWhat stands out3
ClaudeSKILL.mdClaude Code, CoworkNo-code authoring in Cowork
MicrosoftSKILL.mdCopilot Cowork, Copilot StudioCopilot Cowork built on Anthropic’s Cowork technology
OpenAISKILL.mdChatGPT Enterprise, CodexEnterprise skill sharing
GoogleSKILL.mdGemini CLI, Gemini EnterpriseGems moving to skills
GitHubSKILL.mdCopilot, Copilot coding agentRepository-native skills

Pricing and packaging shift frequently; verify current terms before committing. The platform matters less than the operating decision: pick two workflows with high repetition and visible quality variance, baseline them, build the first skills, and stand up the registry and ownership model at skill number one, when it costs nothing.

Skills are the first practical answer to a problem every AI program has: how to turn individual productivity into organizational capability. They are cheap to author, portable across the platforms your teams already use, and powerful precisely because they carry the one thing no vendor can ship: how your firm works.

Name the thing, encode your judgment, govern it lightly, and let the library compound. The tools will keep changing. What you encode is yours.

Notes

  1. Anthropic, "Agent Skills (specification and authoring guidance)", Anthropic, 2026. ↩
  2. Microsoft; Google, "Declarative agent skills and Copilot Studio skills documentation; Skills in the Gemini app and Google Workspace", Microsoft Learn, accessed October 19, 2026; Google Workspace Updates, updated October 13, 2026, 2026. ↩
  3. Charles Lamanna, Microsoft, "Copilot Cowork: A new way of getting work done", Microsoft 365 Blog, 2026. ↩