Watch an operator delegate marketing work to an AI assistant for the first time and you'll see a specific arc. Week one, astonishment: it wrote a campaign brief in forty seconds. Week three, frustration: the fourth brief is structured differently from the first three, the positioning section keeps wandering, and the operator is now writing longer and longer prompts trying to pin the thing down. Week five, the verdict: "it's inconsistent."
The delegation is what broke. Week five reveals that the operator has been re-explaining the job from scratch every single time — and re-explaining it slightly differently, because humans do. The assistant has been faithfully producing what each prompt described: a different job every day.
Nobody runs a team this way. When you hire a person, you don't re-derive how your company writes a campaign brief in every morning's conversation. You hand them the procedure: the template, the steps, what a finished one looks like, what always gets checked before it ships. The whole point of a procedure is that the quality lives in the document, so it stops depending on how well anyone explains the job today.
Agents can be handed procedures too. In the AI-tools world these are called skills, and understanding them is the difference between owning a capable assistant and renting a slot machine.
What a skill actually is
A skill is a written procedure an agent loads and follows: what the deliverable is, what inputs it needs, what steps produce it, what good looks like, what to refuse to invent. Plain files, mostly prose. When you ask for the deliverable, the agent finds the matching skill and executes the method instead of improvising one.
The shift this produces is exactly the shift a good SOP produces in a human team. The tenth campaign brief has the same bones as the first. The definition of done is stated, so it gets checked. And when the output is wrong, you fix the procedure (once) instead of hoping tomorrow's phrasing lands better. Your delegation stops being an oral tradition.
We learned this by doing it at volume. Jinn is an AI company that builds brand infrastructure, and we published our marketing procedures as an open catalog — it's called jinn-skills, it's open source under a license that lets anyone copy and adapt it, and it holds fifty-nine skills, each producing a complete deliverable: a positioning brief, a campaign brief, a launch playbook, ad-copy variants, an AI-visibility audit, a competitor positioning map. Free, readable by anyone, and usable with the common AI assistants. I'm going to use what building it taught us as the lesson plan, because the three hard-won rules apply to any procedure you'd write for an agent — including your own, in a private folder, never touching our catalog.
Anatomy of one skill
Abstractions about procedures are cheap, so open a real one. The campaign-brief skill in our catalog is a single file, about 1,400 words of plain prose, and it has five parts. Every skill you write yourself should have the same five.
The first lines say what the skill produces and when it applies: a campaign brief with objective, audience, single-minded message, channels, hooks, and a success metric, for use "when planning a campaign, launch push, promotion, or content sprint and you need one page everyone builds against." That sentence does the routing. More on why it matters below.
Before any instructions, the file shows the finished artifact as a six-line template: objective, target audience, single-minded message, channels, hooks, success metric. The agent knows what done looks like before it starts, which is exactly the thing a prompt never tells it.
Six numbered steps, and the order is the method: objective and message get locked first because everything else serves them. The steps carry the tests a good strategist would apply. If the user names several objectives, ask which they'd keep if they could only have one; the rest are hopes. If the message can be split into two claims joined by "and," you have two messages, so choose. A success metric is one number, a target, and a window, and if it can move while the objective doesn't, it's vanity and gets replaced. None of that is model behavior. It's the file.
The skill states its inputs: campaign purpose and audience, asked for if missing. When a brand's record is connected, it maps specific fields onto specific brief lines: audience tribes to the target, pain points to the pain you lead with, the top messaging pillar to the single-minded message, tonal attributes to the hook voice, banned words to the copy constraints. And it draws a line it will not cross: "Competitor angles stay yours." No competitor data comes over the connection, so any versus-the-alternative framing has to come from the operator's own market knowledge, never from a field the agent quietly invented.
Expired access, wrong brand, no account at all: the file names each case, says what to do, and in every one the brief still ships in full from the standalone method.
Read those five parts back and notice what you're holding: a marketing SOP that happens to be executable. Nothing in it asks the model to be clever. It asks the model to follow a file that already is.
Rule one: a skill must stand alone
Every skill in the catalog is required to produce a complete, client-ready deliverable from its own procedure — before it's connected to anything. The procedure itself has to carry the method: how to think about positioning, what a campaign brief must contain, what makes ad copy testable.
This sounds obvious and gets violated constantly. The temptation, especially for a company like ours with a brand-data product to sell, is to write thin skills that only work when connected — a procedure with a hole in the middle shaped like your product. We held the opposite line: standalone output must be genuinely good, and connecting your brand's record makes it yours rather than making it work. The catalog's own introduction states the difference in a sentence I'll quote directly: "A positioning brief from the skill's own method is sharp and defensible; one written against a brand's real competitive wedge and enemy is unmistakably theirs."
If a vendor's skills collapse without the vendor's data, you're not looking at procedures. You're looking at a funnel.
That's the honest division of labor: the procedure supplies competence, the brand record supplies specificity. If a vendor's skills collapse without the vendor's data, you're not looking at procedures. You're looking at a funnel.
Rule two: the agent has to pick the right procedure, and you can measure whether it does
The failure mode nobody warns you about: you write twenty procedures, ask for "a teardown of my competitor's pricing page," and the agent runs the generic content skill instead of the pricing-teardown one. A skill that doesn't get picked when it should might as well not exist.
We test this on our own catalog, and published the test itself. One hundred seventy-five realistic marketing requests, shown to three different Claude models, three times each (1,575 trials), where the models see only the catalog's names and descriptions, the same view an agent gets. In the latest published run, the correct skill ranked first in 97.3% of the answers that came back in readable form, across the whole fifty-nine-skill catalog; the 452 answers that came back garbled are left out of that percentage and reported in full alongside it. And one honesty note the catalog itself insists on: this measures whether the right skill gets picked, not how good its output is.
The transferable lesson is bigger than our number. Whether the right procedure gets picked depends on its name and description, which means the unglamorous work of describing each procedure (what it's for, when it applies, when it doesn't) is load-bearing. If you build a skill library and delegation feels random, your procedures probably aren't bad. They're indistinguishable.
If you build a skill library and delegation feels random, your procedures probably aren't bad. They're indistinguishable.
Look again at the campaign-brief description. It names the deliverable, lists its parts, then spends half its length on when: a campaign, a launch push, a promotion, a content sprint. That second half separates it from its neighbors. The catalog also holds a product-launch playbook and a launch-positioning brief, and an agent reading only titles could confuse all three. The descriptions are where they stop overlapping: the playbook is phases and sequencing, the positioning brief is wedge and enemy, the campaign brief is one message and one metric. Write descriptions as if the reader has twenty similar files open and ten seconds to choose. That is the reader.
Rule three: write yours this week
You need no vendor for any of this. The recipe:
For most consumer-brand operators that's five to ten: the launch brief, the email sequence, the ad variants, the competitor check, the weekly content plan.
The inputs it needs, the steps in order, a definition of done, and one example of a good finished artifact.
What this procedure must never invent — the reviews you don't have, the stats you can't back.
Describe it in one unambiguous line, and keep the folder where your assistant works.
That never-invent line deserves more than a sentence, because it's where most agent output goes wrong and where a procedure earns its keep. Our catalog handles it two ways you can copy. Some skills tag every claim: the competitor-profiler labels each line Sourced, Inference, or Unconfirmed, so a reader can see which parts are known and which are the model's guess. Others refuse the job when the inputs can't support it: the programmatic-SEO planner is written to reject a thin keyword-swap page set rather than produce one, because a plausible plan for bad pages is worse than no plan. Pick whichever fits the deliverable.
Then stop prompting the job and start requesting the deliverable.
The switch, in practice.
| The prompt version | The procedure version | |
|---|---|---|
| Write me a campaign brief for the spring launch, we want awareness and signups and some revenue, make it good. | You get three objectives, a paragraph of message, and channels listed because they exist. | The same request with the campaign-brief skill in the folder: the agent asks which objective you'd keep if you could only have one, writes a single message and checks it can't be split, gives each channel a reason, and ends on one number with a target and a window. |
Same model, same operator, same afternoon. The difference is what was written down beforehand.
That folder becomes something your brand has never had before: marketing judgment that persists between conversations, improves when you edit it, and doesn't walk out the door with the person who held it in their head. The models will keep changing underneath it. The procedures are yours.
Questions people ask
Isn't a skill just a long prompt saved to a file?
The file format is the same; the content isn't. A saved prompt describes the job. A skill describes the deliverable's shape, the steps in order, the tests each step applies, the inputs it needs, and the things it won't invent. The tell is whether you could hand the file to a new human hire and get a usable first draft back. If yes, it's a procedure.