Skill

An AI skill is a reusable method for a kind of work. A prompt asks for one result. A tool gives the AI an action. A skill says what to collect, which rules to follow, what resources to use, and how to check the result. The open Agent Skills format requires a SKILL.md file with name and description metadata, and it can bundle instructions, scripts, references, and templates. Agents can load that material in three layers, from a short catalog entry to full instructions and then supporting files. For a service firm, the value is repeatable work. A proposal skill can carry the same intake questions, house template, approval rules, and checks into every proposal. It cannot settle a method the firm has never agreed on.

What does an AI skill contain?

A skill puts the instructions for one kind of work in one place so an agent can find and reuse them. In the open Agent Skills specification, each skill is a folder with a required SKILL.md file. That file starts with a name and description, followed by the work instructions. The folder can also contain scripts, reference material, templates, and other assets.

The structure keeps every company rule out of every request. The Agent Skills overview describes three levels for loading the material. The agent sees each skill’s name and description first. It reads the full instructions when the task matches. It opens supporting files only when the job needs them.

Take proposal writing. The main instructions can set the sequence. Collect the client’s problem. Confirm the scope. Choose the correct template. Draft the proposal. Run the approval checks. A reference file can hold current service descriptions. A script can check for dates, owners, and required clauses. A template can carry the approved layout.

The skill is the package around the method. The method still has to be sound. If scope lives in one manager’s memory and three partners use different approval rules, putting those fragments in a folder does not create a standard.

How is a skill different from a prompt, tool, or workflow?

These four parts solve different problems.

Part	What it provides	Service-firm example	Main limit
Prompt	Instructions for one request	“Turn these notes into a project brief”	Must be supplied again
Tool	An action the AI can take	Read the CRM or create a document	Does not explain the firm’s method
Skill	Reusable know-how for one kind of work	Create and check project briefs	Needs a stable standard and owner
Workflow	The ordered route from trigger to record	New call ends, brief is created, reviewed, and filed	Coordinates the job rather than teaching every step

A prompt can be part of a skill. A tool can be called by a skill. A skill can run inside an AI workflow. The boundaries matter because each part fails in a different place.

If the CRM tool returns the wrong record, fix the connection or query. If the skill puts risks in the wrong section, fix the instructions, example, or check. If the finished brief never reaches the project manager, fix the workflow handoff. Calling all three problems “the AI” hides the owner of the repair.

A skill also does not replace the workflow. It can teach an agent how to complete a step. The workflow still needs a trigger, context, handoff, and record. A person may still own the decision at the end.

When should a firm turn work into a skill?

Create a skill when the same work repeats and the best version of that work can be written down. Four conditions make the candidate clear.

The task repeats. A one-off board memo can stay a prompt. A project brief created after every discovery call is a candidate for a skill.

The standard is stable. The people responsible for delivery should agree on required inputs, acceptable output, and the failure route. A skill cannot settle a live management dispute.

Real examples exist. Include strong outputs, common edge cases, and known mistakes. Generic examples teach generic work.

Someone owns it. One person or team approves changes, watches failures, and removes stale rules. Shared use without a named owner turns a skill into an old procedure that happens to run.

The usual failure starts small. One employee keeps a prompt in a notes app. Another copies it and changes two lines. A third starts from last month’s output. The same missing field gets repaired by hand, but nobody puts the fix back into the shared method. The team is working around a gap in its documented process.

Start by writing the work itself. Document the process first, including judgment points and exceptions. Then package the stable part as a skill. The order matters. A skill makes instructions easier to repeat. It does not make those instructions true.

How do you test and maintain an AI skill?

Test the work the skill produces. Do not use the agent’s statement that it followed the instructions as the test.

Anthropic’s October 16, 2025 engineering guide says to start with evaluation. Run representative tasks. Observe specific failures. Build the skill around the missing knowledge or procedure. This keeps the files tied to a real production gap.

Build a small test set from work the firm already handles. For a proposal skill, include a clean standard request, incomplete notes, conflicting dates, an unsupported promise, and a request outside the service scope. Score whether the skill collects missing information, uses the right template, keeps claims inside the evidence, and routes exceptions to the right person.

Use a written eval so the score survives a model change. Run the same cases before and after each skill edit. Record the skill version with each result. A change that fixes one missing field but creates two false promises is a regression, even if the new proposal reads better.

Maintenance follows the delivery standard. When the firm changes a service, template, approval rule, or tool, the skill has changed even if nobody edited the folder yet. The owner updates it, reruns the test set, and records the new version. That is ordinary process ownership applied to agent work.

What can go wrong with an AI skill?

The first risk is making a bad process repeatable. A detailed skill can look mature while carrying the wrong approval rule into every job. More instructions can make that error harder to notice.

The second risk is scope. A single “do everything” skill loads rules that do not belong together. Proposal writing, contract review, invoicing, and client onboarding have different owners and permissions. Split them along real responsibility lines. Smaller skills are easier to trigger, test, and retire.

The third risk is access. Skills may contain executable code and instructions to use the network. Anthropic recommends installing skills only from trusted sources and auditing bundled files, dependencies, and external connections before use. A downloaded skill should be handled like downloaded software.

Give each skill the least access its task needs. A proposal skill may read approved service information and create a draft. It does not need permission to send the proposal, change the CRM owner, or read payroll. The decision about autonomy still depends on whether the result can be checked and whether the action can be undone.

An outdated skill creates a quieter risk. The file may still run while the service, template, or approval rule has changed. A named owner and repeatable test are what make that drift visible.

Where does a skill sit on the Delivery Model Ladder?

A production skill belongs at Stage 2, Augmented, on the Delivery Model Ladder. At that stage, AI is part of a named workflow with a trigger, current context, tool access, a review point, and a final record. The skill supplies reusable know-how inside that route.

The same file can sit at Stage 1, Enhanced, when a person manually chooses it, pastes in the material, checks everything, and moves the output by hand. Packaging instructions does not change the delivery model by itself. The surrounding route decides the stage.

This is also the connection to an AI employee. The role needs a named scope of work. Skills carry part of the method for that scope. The workflow supplies the start, handoffs, limits, and record. A person still owns the outcome until the system earns more autonomy through measured work.

Quick answers

Is an AI skill just a long prompt? No. A skill can include prompts, but it is discovered and reused across matching tasks. It can also carry scripts, references, templates, examples, and rules for when to load them.

Does a skill give an agent new tools? A skill can teach an agent how and when to use tools, and it can bundle scripts. The tool provides an action. The skill provides the method around that action.

Who should own an AI skill? The person accountable for the delivery process should own its standard. A technical teammate may maintain the files, but process changes need approval from the person who answers for the work.

Are downloaded AI skills safe? Treat them as software. Read the instructions, scripts, dependencies, permissions, and network calls before use. Run them somewhere with limited access first, and test representative work before granting more access to client data or production systems.