The prompt is no longer the right unit of work with Claude. The right unit is the Skill. A Claude Skill is a folder holding a SKILL.md file, your procedure written down once, and the templates that go with it. Claude loads it on its own when it spots a matching task. You no longer paste four hundred words of context at the start of every conversation.
This shift has a date. Anthropic announced Agent Skills on 16 October 2025, then published the format as an open standard in December 2025. Since then, the skill that matters is not phrasing. It is the ability to describe your own process precisely enough for an artificial intelligence to run it without you.
I made this point in a LinkedIn post in March 2026 that drew around 600 reactions. What I see on client work since then confirms it. The teams making progress are not writing better requests. They have put their processes in the right place.
What exactly is a Claude Skill?
A Claude Skill is a folder the model can read, whose only mandatory file is called SKILL.md. That file opens with a YAML header carrying two fields, a name and a description. The name is capped at 64 characters, lowercase and without spaces. The description can run up to 1,024 characters.
The rest of the folder is up to you. Reference markdown files, deliverable templates, executable scripts, database schemas. Below the header comes the procedure itself, written like an onboarding guide meant for a new hire.
The most underrated field is the description. It is what Claude matches your request against when deciding whether to load the Skill or skip it. Anthropic's documentation, checked on 23 August 2026, insists on this. A description that says what the Skill does without saying when to use it produces a Skill that never fires.
Tandem has been running its own Skills internally since spring 2026. Sales proposals, non-disclosure agreements, meeting notes, case studies, end-of-project reports. Every recurring process replaces a playbook that had to be pasted back into each conversation by hand.
How Claude loads a Skill in three layers
Claude never reads a Skill in full at once. It discovers it in three stages, what Anthropic calls progressive disclosure. At startup, only the metadata enters the context, for roughly 100 tokens per Skill. That is the cost of a sleeping Skill.
When your request matches the description, Claude reads the body of SKILL.md. The documentation puts that layer under 5,000 tokens. Bundled resources cost nothing until someone opens them. A script that runs sends only its output into the context, never its code.
This economy is the opposite of the one connectors follow. An enabled connector is reread at every message and weighs on the window permanently. I put numbers on that snowball effect in our article on how prompting changed in 2026. A Skill is only paid for at the moment it earns its place.
Why the prompt is no longer the right unit of work
A prompt settles one task. It lives inside a conversation, then disappears with it. The colleague next door starts over from scratch, and a rule that changes has to be rewritten in ten different places. That is the design flaw of the prompt library.
A Skill settles the task and every task that resembles it. You edit SKILL.md, and the new rule applies everywhere the Skill fires. The process stops being an individual habit. It becomes a company asset.
Anthropic set out this shift in an engineering post published on 29 September 2025. The useful discipline is no longer prompt engineering but context engineering, the craft of choosing what the model has in front of it when it answers. The team also describes context rot, the way recall degrades as the window fills up.
I put it differently in a July 2026 post on the end of prompting. You are no longer talking to a model on its own. You are steering a set of things, made of conversations, Projects, Skills and connectors.
Projects, Skills and the context folder, which carries what
Three mechanisms look alike and do not serve the same purpose. A Claude Project carries reference documents. A Skill carries a procedure. The folder you share with Claude Cowork carries your working environment.
| Building block | What it carries | When it is read | Who can reach it |
|---|---|---|---|
| The prompt | A one-off instruction | The moment you send it | You alone, for the length of the chat |
| The Project | Large reference documents | By retrieval, on every question | The team, on Team and Enterprise plans |
| The Skill | A procedure and its templates | When the description matches the request | Depends on the surface, from laptop to shared repo |
| The Cowork folder | Your working context and files | Every session opened on the folder | You, on your own computer |
A Claude Project takes heavy files. When the knowledge base outgrows the context window, Claude switches to retrieval and reads only the passages that serve the question asked. A Skill works the other way round. It stays small, and its description is what triggers the read.
The context folder is still the highest-return brick of the three. Three two-page markdown files do more work than fifty carefully written prompts. I detailed the exact structure in our practical guide to Claude Cowork, and the full path in how to master Claude in ten steps.
Three Skill patterns that hold in production
Three shapes of Skills keep coming back in our deployments. You pick between them based on how uncertain the task is, not on the team's line of business.
- The generator. A main Skill leans on sub-Skills that frame the output. It is the right pick to standardise a deliverable with distinct parts, such as a technical bid in construction.
- The inversion. The Skill asks every critical question before it executes. Nothing is assumed, no gap in the brief. It is the right pick when a wrong assumption is expensive.
- The chained inversion. The Skill alternates question and action, step by step. It is the right pick for a project that only takes shape as it goes.
I published this grid in a one-pager on Skills in April 2026. It still holds, and the inversion pattern is the one that changes the most inside companies. A Skill that asks before it acts produces far fewer deliverables you have to throw away.
What Rakuten and Anthropic's finance team do with their Skills
Rakuten is the first named case Anthropic put forward when it announced Skills in October 2025. The Japanese group equipped its management accounting. Claude processes several spreadsheets, catches anomalies and produces the report following in-house procedures. A task that used to take a full day now fits in an hour. Box, Canva and Notion appear in the same announcement.
The most telling example comes from the vendor itself. Anthropic's finance team runs roughly 150 shared Skills, stored in a version-controlled Git repository, according to an interview published by CFO Connect in June 2026. Tim Ross, Finance AI Product Lead, and Lisa To, who heads finance systems, describe validating a full year of historical data in about twenty seconds.
The same interview recalls a figure that puts things back in proportion. Gartner estimated in 2025 that 63 % of AI initiatives in finance fall short of their goals. Skills do not fix that on their own. They fix one precise cause, the fact that process knowledge stays inside people's heads.
This is exactly the work Tandem does with its small and mid-sized clients. Across more than 40 projects, the hard part has almost never been the model. It has been getting the process written down. Our AI audit exists first to spot the processes that deserve a Skill, and the ones that do not.
Should you write your Skills or record them?
Since July 2026 there is a third route. Claude Cowork's recording feature captures your screen while you do the task once, then Claude derives a reusable Skill from it. It lives in the desktop app, on the Pro, Max and Team plans.
That feature solves the real point of friction. What blocks teams is not SKILL.md syntax, it is having to explain in writing how they work. Showing something once takes less effort than drafting it.
The limits to know before rolling out Skills in a company
The limit that really blocks people is not technical. Custom Skills on claude.ai are individual. Each person uploads their own, and no admin console pushes them to the organisation. An IT department that wants a shared repository has to go through the API or through Claude Code.
Skills do not sync from one surface to another. A Skill uploaded to claude.ai does not exist in the API, and the reverse holds too. In Claude Code they live in a folder on the machine or in the code repository. So pick your reference surface before you produce dozens of files.
On the API, Skills run in a sandboxed container with no network access and no package installation at runtime. A Skill that needs to call an external service will not work there. In Claude Code network access is full, which opens far more doors and demands far more care. Our guide to Claude Code for non-technical profiles covers that surface.
Should you replace your prompts with Claude Skills?
Yes, as soon as a task comes back more than two or three times a month and its output format is stable. That is the threshold past which re-explaining the context costs more than writing it down once. Meeting notes, sales proposals, recurring reports, compliance checks, all of that deserves a Skill.
No for exploration. A one-off question, a draft, a single analysis are all handled well by a good prompt. The prompt is not dead. It has simply stopped being the place where know-how compounds.
A good start is two Skills, not twenty. Take the most repeated process in your team, write it down, let Claude replay it three times, fix the file between each run. By the third pass you will know whether the format holds up.
Only then does the question become collective. Who writes the Skills, who signs them off, where they live. That is the substance of our Claude training sessions for teams, and it is also the real subject behind the difference between AI agents and AI workflows.
