AI agent skills for comms teams: what they are and how to adopt one safely

Checked against each vendor’s own documentation on 9 November 2026. Skills are changing quickly, so I re-check this guide every three months, and sooner when a vendor changes its terms; the Briefing says when it changes. Nobody named here paid to be included.

The way Comms With AI approaches skills has three layers, and each depends on the one beneath it.

Most writing about skills covers only the middle layer. For a comms team, the other two decide whether a skill is safe to use.

An AI agent skill is a reusable set of instructions for a particular task, often stored in a file called SKILL.md. Alongside it can sit reference material, templates, examples and scripts. Compatible tools load a skill automatically when a task matches its description, or let you call it by name. The format is an open standard, Agent Skills, first published by Anthropic.

The tool reads only each skill’s name and description until a task needs it, then loads the rest. That is why a skill can carry far more than a prompt you paste in: it takes no room until it is used.

Here is a comms example: a skill that checks a draft statement against your Position Register. Choose a note to see which lines it describes.

With the skill installed, the assistant follows the same check whenever the task comes up, and nobody has to paste the instructions or remember every step. The dependency is plain too: the skill needs a current, trustworthy register to check against.

Where skills run today, from each vendor’s documentation:

One caution: each tool still handles scripts, connectors and behaviour in its own way, so a skill moved between tools needs testing in each one.

A skill instructs the tool, and it works with the access the tool already has: the files it can read, the connectors it has been given, the model behind it. Where the tool can reach your shared drive, so can the skill.

A skill file alone cannot guarantee compliance with your AI policy. Someone can open a fresh chat, skip the skill, or give it a task it was never written for. Admin controls on business plans help, by deciding who can install, share and publish skills. A dependable arrangement adds approved tools, access controls, training for the people using them, a check on what they produce, and a named owner. That is governance with a file at its centre.

When an AI agent creates or sends communications to employees, customers, journalists or other stakeholders, it speaks for the organisation. A press office assistant, a bot answering staff questions about a change programme, a tool drafting replies to stakeholders: what each says, in what tone, and what it declines to say are communications decisions, whoever builds it.

So the communications function should be involved in setting the content standards, the approved positions and the review arrangements. That material lives in three places. Foundations hold the organisation’s profile and rules. A position register records what you have said and will stand behind. Tailored skills carry both into the tool. The free Position Register template is a good place to start.

Treat a third-party skill, plugin or agent pack like any other tool that will act in your name. Six steps, set out in full in the Governed Install Checklist:

1. 01

   Verify the source

   Name the publisher and take the files from a source you trust, before you assess anything else.

2. 02

   Set the risk tier

   Instructions only, reads your information, or can act: the tier decides who signs off.

3. 03

   Pin it and read it

   Record the version, switch off automatic updates, and read every file, scripts and connectors included.

4. 04

   Test it safely

   Run it on non-sensitive material first, and have someone who knows the work check the output.

5. 05

   Count the real cost

   One-off set-up, then model usage, subscriptions, paid data and the staff time to maintain it each month.

6. 06

   Name an owner and log it

   Record it in your AI Tool Register, with a review date and a named person who can switch it off.

Two free skills are being tested now, one to check a draft against a position register and one to reconcile conflicting feedback; this guide will link them when they pass. The light-touch version is already on the site: your Foundations profile holds your organisation’s rules, and Plan and Message Test apply them to your work. For a team that wants rules, structures and skills set up in its own tools, tailored skills can be scoped through Deploy, where most teams invest £1,500 to £7,500, and kept current through Manage, from £300 a month. That work is done in conversation with your team, and Consult, at £300, is the lighter first step.