Ten Rules for Skills: Anthropic Turns Capability Into a Governed Platform

01 THE RULESET ARRIVES

Anthropic's publication of ten rules for building Claude Skills is easy to read as documentation hygiene, and that reading undersells it. A roughly 12-minute video from the channel Jay E | RoboNuggets, observed at approximately 124,800 views via yt-dlp on 2026-11-07, walks through the ruleset and its practical implications. The measured facts are modest — a short explainer video with a substantial audience. The signal is in the genre: when a platform publisher writes rules for how a capability must be named, secured and distributed, it is treating that capability as a platform primitive rather than a workaround.

Platforms govern through conventions long before they govern through enforcement, because conventions decide what is easy, discoverable and trusted. A ruleset is the cheapest governance instrument available: it shapes builder behavior without a single line of enforcement code. The remaining sections read the ten rules as an instrument — what they standardize, what they quietly centralize, and where their authority stops.

02 WHAT A SKILL ACTUALLY IS

A skill is a packaged unit of procedural knowledge. Its core is a manifest file, SKILL.md, whose frontmatter carries a name and a description; its body holds instructions for how to perform a task. Around that core a skill may bundle optional scripts and resource files. Nothing in that structure executes on its own — an agent loads the package and decides, from the description, whether the skill is relevant to the request at hand.

The design principle is progressive disclosure. The agent holds only the thin index of available skills — names and descriptions — in its working context; when a request matches, it opens the relevant body; when that body calls for computation, it runs the packaged scripts; resources are pulled in only as referenced. Context is spent on what is used, not on what is installed. That economy is precisely why the description field matters so much: it is the only part of the skill the model sees when deciding relevance, which makes the manifest a discovery surface as much as a documentation page. The anatomy is diagrammed below.

03 READING THE TEN RULES AS A TAXONOMY

The ten rules do not arrive with an official taxonomy, so the grouping below is an editorial reading — a way to see what the rules collectively optimize for. Four themes emerge. Naming and discovery rules govern how skills are titled and described so that an agent can match them to requests reliably. Security and trust rules govern what a skill is allowed to assume, execute and request, and how provenance is communicated. Performance and cost rules govern context economy: what gets loaded, when, and at what token expense. Distribution and compatibility rules govern packaging, versioning and portability across environments.

Read as a set, the themes reveal intent. Discovery rules make the ecosystem navigable; security rules make third-party packages insurable; cost rules protect the context budget that progressive disclosure depends on; distribution rules decide whether skills remain inside one product or become a cross-platform format. The rough weighting of the rules across these themes, charted below, is an interpretation rather than a count endorsed by Anthropic.

04 NAMING IS CURATION POWER

The least glamorous rules carry the most structural weight. Because an agent selects skills by matching descriptions against requests, the person who writes a good description controls when a capability activates. Naming and description conventions are therefore not style guidance; they are the ranking algorithm. A skill whose description over-matches will be selected too often; one that under-describes will never fire. Platform-defined conventions for how descriptions must be written effectively become the search index of the capability ecosystem.

That makes naming a curation power. Whoever sets the vocabulary — canonical verbs, domain terms, anti-patterns to avoid — shapes which skills surface in ambiguous situations, much as app-store search terms shape which applications users find. It also concentrates interpretive authority: a builder whose phrasing aligns with the published conventions wins discovery that an equally capable but idiosyncratically described competitor does not. None of this requires bad intent to matter. The governance question is simply whether the conventions are set openly, documented and contestable — or accrete informally around examples until they function as unreviewable policy.

05 SKILLS AS AN INJECTION SURFACE

A skill is instructions a model will follow, delivered from outside the model's training. That is its power and its risk in the same sentence. A malicious or careless skill is an injection surface: prompt-level content that arrives packaged as expertise, carrying whatever instructions its author embedded. Unlike a transient prompt injection, a skill persists in the environment, waits for a matching request, and arrives with the credibility of an installed tool.

The security rules in the ruleset read as responses to that surface: treat third-party skills as untrusted input, keep their permissions minimal, prefer scoped execution to ambient authority, and make provenance visible. The underlying principle is least privilege, applied to procedural knowledge — a skill that summarizes spreadsheets has no reason to invoke network access or to instruct the agent to skip its usual confirmation behavior. For enterprises, the practical consequence is a supply-chain discipline for skills analogous to dependency management in software: inventories, review and signing are the difference between a capability marketplace and an attack vector with a catalog.

06 THE COMPETITIVE FRAME

Skills sit beside an existing interoperability layer rather than replacing it. The Model Context Protocol, Anthropic's open standard for connecting models to external tools and data, standardizes connectivity; skills package procedure — the knowledge of how to accomplish a task well. The two are complementary in principle, and the ruleset's compatibility provisions will determine in practice whether skills remain product-specific or harden into a portable format other agent builders adopt.

The competitive stakes are visible in rivals' responses, including Google's skills-oriented moves for Gemini. Platform primitives compete on defaults: whoever defines the manifest format, the discovery conventions and the trust model first gets to define them for everyone else, because ecosystems standardize on the earliest adequate specification. Publishing a ruleset is part of that race — it converts a vendor capability into a candidate standard with documented expectations. The contest is less about which agent executes skills better than whose conventions the ecosystem's skills are written for.

07 WHAT TO WATCH NEXT

Four observable signals will indicate where the platform is heading. First, enforcement: whether the rules remain published guidance or acquire mechanical validation — manifest checks, submission review, signing — at distribution points. Second, ecosystem growth: the size and diversity of published skills, and whether third-party repositories emerge outside Anthropic's control. Third, portability: whether other agent platforms adopt the SKILL.md conventions, which would mark the format rather than the vendor as the standard. Fourth, the incident pattern: whether the first publicized skill-based injection or exfiltration episode arrives, and how the trust rules change in response.

The measured record so far is thin — a ruleset, an explainer video, an audience. The direction, however, is legible. Capabilities that once lived as prompt folklore are being converted into governed, documented, distributable platform objects, and the ruleset is the moment of conversion. What remains open is whether governance follows the rules into enforcement, or stays, as conventions so often do, advisory until something breaks.