Skills, Plugins and Agents: Building Capability You Can Deploy

Moving from solo prompts to deployable capability

When an individual discovers a prompt structure that summarises a site log or drafts a client email in two minutes, that is a private efficiency. It saves their afternoon, but it leaves the rest of the business exactly where it was. If that person changes roles next month, the operational gain leaves with them.

Building a team asset means packaging that workflow so anyone on the team can use it with consistent quality. In enterprise platforms like Claude and Microsoft 365 Copilot, this packaged capability is referred to as a custom skill, a plugin or an agent. While the technical names differ between vendors, the core operational idea remains identical.

A custom skill is simply a documented business process translated into clear instructions, reference files and boundaries that an AI model can follow. You do not need software engineering experience to direct this work. Capturing the process sequence, the business logic and the quality exceptions is the primary task, and that is standard business analysis work.

The three mechanisms: skills, plugins and agents

To direct technical teams or external consultants with confidence, you need to understand what each term means and what role it plays in your workplace.

Under the Hood: What skills are made of and how they fire

Let us look under the hood to see how a skill actually operates on the file system and inside the model workspace.

At a technical level, a skill is not compiled software code. It is a folder containing a structured instruction file, typically named SKILL.md, accompanied by reference materials such as spreadsheets, templates, checklists or calculation guides. The instruction file contains three primary components: a distinct name, a concise description of what the skill does, and the step-by-step instructions the model must follow.

The description field carries immense operational importance. When a user enters a query in Claude or Copilot, the system evaluates the prompt against the descriptions of all installed skills. The description acts as the semantic trigger. If your description states that the skill reviews structural engineering quotes against regional council standards, the system routes the request to that skill automatically. If the description is vague, the skill fails to activate.

The same fundamental structure runs across leading enterprise platforms. While Claude packages these as custom skills and projects, Microsoft 365 Copilot packages them as declarative agents or Studio skills. In both environments, system administrators can publish a completed skill to the entire organisation or restrict it to defined security groups.

Here is the critical design point to take into any room with IT or software developers: a skill provides instructions and local reference templates, but it does not grant underlying data permissions. A skill cannot see data that the user running it cannot access. If your process requires live commercial margins or confidential client rates, those figures must stay inside a permissioned database that the skill queries using the employee’s existing security credentials, rather than being hardcoded into the skill text.

The engineering firm: a worked design example

Consider how this design operates inside a mid-sized regional engineering consultancy. The firm handles civil infrastructure, geotechnical reviews and commercial structural tenders. Rather than expecting forty engineers to invent their own prompts, the leadership team scopes three focused skills.

First, an estimating skill guides junior engineers through quantity extraction and standard labour formulas based on historic job actuals. Second, a workshop fabrication skill formats bill-of-materials tables directly for the CNC machinery scheduling software. Third, a finance review skill audits completed project accounts against client payment milestones.

The firm separates the estimating capability into two tiers. A junior estimating skill applies standard regional labour cost bands and produces drafts for peer review. A senior estimating skill connects to executive margin approval schedules and commercial risk tables. By mapping the capability to user roles, the firm maintains operational consistency without exposing sensitive commercial thresholds across the wider team.

The five-stage deployment framework

To lead this work across your own business, use this structured framework from initial identification to final release:

1. Understand: Distinguish between instructions that teach a model a process and connectors that link it to external systems.
2. Scope: Choose one high-frequency, text-heavy process with clear source inputs and predictable outputs.
3. Direct: Document the operational rules, required inputs, edge cases and quality criteria before approaching technical leads.
4. Deploy: Work with your administrator to stage the release, testing first with a small pilot group before broader rollout.
5. Govern: Establish human review checkpoints and verify that no confidential data is embedded in the instruction files.

For organisations reviewing their public accountability and procurement obligations, consult the New Zealand responsible AI guidance to align workflow governance with standard administrative expectations.

Try this: scope your first team skill

Identify one recurring reporting or review task that your team completes every week. Draft the procedural brief on a single page by defining four concrete elements: the exact documents required as inputs, the three non-negotiable formatting rules, the specific exclusions that must be flagged for human review, and the expected final output format. Handing this procedural brief to a developer allows them to build a working skill in hours rather than weeks.

Prompt template: skill specification builder

Copy this prompt into your preferred AI workspace to turn a standard operating procedure into a clean skill specification: