Chat is the on-ramp: a plain-English guide to agents, and when a skill beats a prompt

If your company rolled out Copilot or Claude this year, you have probably settled into a rhythm: open the chat window, paste something in, ask a question, copy the answer back out. It works, and it is genuinely useful. It is also the smallest fraction of what these tools can do, roughly like owning a laptop and using it exclusively as a calculator.

This post is the guide we keep wishing we could hand people. First, what actually changes when you move from chatting with an AI to handing work to an agent, in plain English with no demo-video hand-waving. Then the question that comes right after: when should you write a skill, and when should you just ask?

The state of play, in three numbers

If the calculator line sounds harsh, the research says it is roughly where most organisations are:

• 62% of organisations are at least experimenting with AI agents, per McKinsey’s State of AI survey. But the gains cluster with the few who changed how the work is done: high performers are nearly three times as likely to have fundamentally redesigned workflows around AI, and that redesign is one of the strongest predictors of real business impact in the whole survey.
• Only 16% of AI users qualify as what Microsoft calls Frontier Professionals, the people who routinely use agents for multi-step workflows, in the 2026 Work Trend Index, built on surveys of 20,000 workers across ten countries plus trillions of anonymised productivity signals. The other 84% are, statistically speaking, still in the chat window.
• The gap shows up in output. In the same Microsoft research, 58% of AI users say they now produce work they could not have produced a year ago. Among those frontier professionals, it is 80%.

So: nearly everyone has the tool, about a sixth of users have found the next gear, and the difference is visible in what they ship. The rest of this post is about that gear.

The difference between chatting and handing over

A chat assistant answers you. An agent works for you. The practical difference comes down to three words, each worth spelling out:

• Integrated. An agent like Claude Code runs on your machine and connects to the systems you already use: your file folders, your email, your calendar, your finance tools, whatever your team exposes through connectors. The plumbing standard is called MCP, and you need to understand it about as well as you understand how email finds your inbox. The point is: no more pasting.
• Contextual. Because it can open the actual folder, it works from the real thing: all forty contracts, the actual export with the broken column, the brand guidelines PDF nobody reads. You stop being the courier who shuttles context in and out of a chat box one screenful at a time.
• Continuous. It does multi-step jobs: read these, cross-check that, draft this, flag anything odd, then reports back when it is done. You review an outcome instead of supervising every step.

Despite the name, Claude Code is not just for programmers. It is a general-purpose agent that happens to be excellent with files. If your work lives in folders full of contracts, invoices, decks and exports, you are exactly who it is for. The same idea ships in friendlier clothes as Cowork in the Claude desktop app. Either way, the shift is the same: from asking questions to delegating jobs.

When you should NOT write a skill

Now the part most guides skip. A lot of tasks need no setup at all: the model is already good at them, and the fastest path is to just hand them over:

• “Summarise this sixty-page report and pull out anything about pricing.”
• “Something looks off in this export. Find it.”
• “Draft a first reply to this complaint. Firm but warm.”
• “Rename everything in this folder to date-first.”

What these have in common: they are one-offs, you can describe them fully in a sentence, and you have no house rules about how they should be done. A useful inversion: if the task would not justify writing instructions for a new starter, it does not justify a skill. Ask, take the answer, move on.

When a skill starts paying rent

A skill is nothing exotic. It is a short document of instructions the agent reads before doing a job: your checklist, your standards, your edge cases, written down once. (Under the hood it is literally a folder with a `SKILL.md` file in it.) The tell that a task deserves one is simple: would you correct the AI’s output the same way twice? If yes, that correction belongs in a skill, not in your patience.

• The task repeats: weekly, monthly, every new client.
• There is a house standard the output must meet, and “close enough” is not.
• The first draft is always 80% right and wrong in the same ways.
• You would have to train a new hire to do it.
• More than one person does it, each slightly differently.

Three examples along the spectrum

Contract review. Ask with no skill and you get a competent, generic review: the internet’s average opinion of what matters in a contract. A skill changes what “review” means: the liability cap you will not go below, the two indemnity patterns that always go to legal, the fallback wording you have pre-approved, the definition of “done” being a filled-in risk table rather than prose. Now every contract gets reviewed against your playbook, including the ones reviewed at 4:55pm on a Friday.

Weekly invoicing. The purest case for a skill: where the time exports live, the rates table, the rounding rule, how the reference number is built, the subject-line format, and the check that hours reconcile against the tracker before any draft goes out. That is ten steps you currently do on autopilot, which is exactly what a skill is for. The task collapses to the same sentence every Friday: “run the weekly invoicing,” then you review the drafts. The judgment stays yours; the choreography stops being.

Designing to a standard. Brand colours, tone of voice, which slide layouts are allowed, the words your style guide bans. Without a skill, “make it on-brand” is a review cycle: someone drafts, someone else corrects. With one, the standard is a starting condition instead of a correction round: the first draft already uses the right palette, the right voice, the approved disclaimer block.

And plenty stays skill-less forever: brainstorms, one-off analyses, naming things. The spectrum is the point: the goal is not to skill everything, it is to notice which of your tasks have quietly earned one.

Now imagine the whole team runs the same skill

Here is where it compounds, and where, in practice, it breaks. The moment a skill is useful it starts to spread: you paste it into a chat thread, three teammates save three copies, someone tweaks theirs. Then legal updates the liability threshold in the contract-review skill. Your copy knows. Theirs does not, and it will keep confidently reviewing contracts against last quarter’s rules. Version drift is the oldest problem in operations, now in a new file format.

This is the problem shareskills exists to solve. Your team’s skills live in one library: publish once, teammates install by asking their agent, and every run starts from the current version instead of a copy from whenever someone first heard about it. Roles carry through: anyone can propose an improvement, maintainers approve it, and risky changes are gated for human review before they reach anyone’s machine. If you want to see the shape of it in practice, we wrote up how one team runs a whole meeting on six shared skills.

The progression, then: chat got your organisation through the door. Agents do the job where the job actually lives: integrated, contextual, continuous. Skills make the output yours instead of generic. And a shared library makes it your team’s instead of yours. Start small: pick the one task you do every week that you would have to explain to a new starter. That is your first skill.