Platform skills — the ones every agent already has

Principle — platform skills are methods, tenant skills are your business. The shared library holds general ways of working (research a competitor, read a regulation, size up a lead). Anything specific to your industry or your data belongs in your own skills. More in Design principles.

Every agent starts with a set of platform skills it did not have to be given. They are general working methods, they load on demand like any other skill, and you do not attach them one by one.

Why they aren't just "more skills on the agent"

A library of a few hundred methods cannot go into the prompt. Listing them all measurably breaks the choice: with 207 skills listed in full, the model picked the right one 30% of the time; with the list first narrowed to the 15 most relevant, 51%. So the platform retrieves before it lists.

Retrieval runs on this turn's message, over both libraries at once — the shared one and yours. Your own skills are not disadvantaged by sharing the space with general ones: measured on 18 genuinely industry-specific skills mixed into 193 general ones, the industry skills were recognised more reliably (86% vs 79%), because industry vocabulary sits far away from general method vocabulary.

Below roughly eight skills nothing is retrieved — the full list is simply shown. At that size the list is already accurate, and retrieval could only introduce a new failure: the right skill missing from the shortlist, with the model then answering from its own knowledge and no sign in the log.

What you do and don't control

Write them in English

Platform skills are written in English — description, instructions, menu_label and the generated trigger phrases.

This is not only about demos. Measured on the same library of 193 Chinese skills, switching the three platform skills from Chinese to English raised how reliably each was recognised: competitor research 75% → 83%, policy reading 83% → 100%, lead qualification 50% → 62%. An English skill sits further away, in vector space, from a crowd of Chinese ones, so it is easier to pick out. (The advantage comes from being different, not from English as such — in an all-English library it disappears.)

Language does not affect whether the model can call the skill: both a self-hosted 35B model and a hosted one emitted correct, well-formed load_skill calls from Chinese and English descriptions alike. Choosing English is a product decision, not a technical workaround.

Your own skills are a different matter — write them in whatever language your customers speak.

Have the platform research before the skill runs

A skill can ask for its material to be gathered before it starts. Put this line at the top of instructions:

PREFETCH: web x4

The platform turns the visitor's message into that many search queries, runs them, and hands the skill a numbered table of facts to cite. PREFETCH: web on its own means the default of two queries; leave the line out and nothing is searched.

Ask for width only where the question has several parts. A product review — pricing, limits, export — measured 2,957 words on five citations when swept narrowly, with every price a specific figure that appeared in no source; at four queries the same question came back with 17 citations. Skills that draft, rewrite or calculate should declare nothing at all: a search they do not need is a straight loss in time and money.

It is not a substitute for writing the rule down, but it is the half that works. The same skill's instructions already said never state a price that is not in the material, and it stated them anyway. Material changed the answer; the sentence did not.

Writing one that actually gets used

The same rules as any skill, plus one that only shows up at scale:

• description says when, in the words a customer would use. It is what retrieval matches against.
• instructions end with what the skill does not cover. With hundreds of skills, near-neighbours are unavoidable; that closing line is the only thing that lets the model notice it opened the wrong one.
• Trigger phrases are generated for you — eight plain-language ways someone might ask for it — and you can edit them. They are matched against, never shown to visitors, and your edits are never overwritten by regeneration.

There is a check worth running after you write one: ask the library which skill each of its own trigger phrases retrieves. If a skill's own phrases keep coming back as other skills, that skill is in the library but will not be reached — merge it with the one that keeps winning, or make the two say plainly how they differ.