Reusable AI work

A skill is a process
your agent can use again.

Good skills capture the steps, examples, decision rules, quality checks, output, and approval boundaries behind work that already succeeds.

The useful definition

A skill is a small operating procedure for AI.

It is more durable than a single prompt and narrower than an entire agent.

1WHEN

When should this skill be used?

2INPUTS

What information, files, or access are required?

3STEPS

What sequence should the agent follow?

4CHECK

How should it validate the result?

5RETURN

What exact output should it produce?

6APPROVE

What must a person approve before anything changes?

A strong skill is a small operating procedure: clear enough to repeat, test, improve, and share.

Do not mix up the pieces

Prompt, skill, app, agent, and routine each have a different job.

Building blockWhat it doesExample
PromptA one-time instruction for the current conversation or task.Ask for a project brief from these notes.
SkillA reusable process with instructions, examples, validation, output rules, and approval boundaries.Create a consistent project brief from approved notes.
App or connectorA connection that lets an AI system read from or act in an outside service.Read a file from Drive or create a draft in Gmail.
AgentA configured worker with a job, context, tools, skills, checks, and channels.Own the weekly project-update workflow.
Routine or scheduleA rule for when a proven workflow should run.Run the project-update skill every Friday morning.

Create your first skill

Start from a task that worked once.

Do not write a theoretical skill first. Complete the task, review the result, then capture the method.

  1. 01

    Complete the task once

    Work through the task manually with AI assistance and keep the inputs, corrections, decisions, and final result.

  2. 02

    Name the repeatable method

    Identify the steps that mattered, the information that was required, and the mistakes the process must prevent.

  3. 03

    Write the skill

    Document when to use it, required inputs, sequence, validation, output format, and approval boundary.

Use a real or sanitized example so the instructions reflect actual decisions and exceptions.
  1. 04

    Test a new example

    Use different details. The skill should ask for missing information, preserve uncertainty, and produce the promised format.

  2. 05

    Improve the failure rules

    Add examples of unacceptable outputs, edge cases, source requirements, and situations where the skill must stop.

  3. 06

    Install, share, or schedule

    Make the skill available to the right agent or team, then add a routine only when the method is reliable.

Only schedule or share the skill after it succeeds on a fresh example and fails safely on incomplete input.

A skill template you can adapt

Weekly account brief

This is a structure, not a finished skill. Replace every placeholder with your actual sources, definitions, and permissions.

NAME: Weekly account brief

WHEN TO USE
Use when preparing the weekly internal account review.

REQUIRED INPUTS
- Current account list
- Approved CRM fields
- Recent support or meeting notes
- The reporting period

STEPS
1. Confirm the reporting period and source freshness.
2. Collect the approved fields for each account.
3. Separate observed facts from inferences.
4. Flag missing or conflicting information.
5. Rank only by the agreed risk definitions.

VALIDATE
- Every factual claim links to a current source.
- No account is marked at risk from a missing field alone.
- Stale data is labeled, not silently reused.

RETURN
A table with account, observed signal, confidence, source, and suggested next question.

APPROVAL
Never contact a customer or update the CRM without human approval.

Quality checklist

A skill should be easier to trust because it is easier to inspect.

✓

The skill states exactly when it should and should not run

✓

Required inputs and missing-input behavior are explicit

✓

Every step is observable enough to debug

✓

The output format is concrete, not just “make it good”

✓

Validation checks happen before the result is returned

✓

Sending, publishing, buying, deleting, or changing live systems requires approval

✓

A fresh test example works without copying the original answer

How current platforms handle skills

The concept is portable. The controls differ.

ChatGPT skills

Reusable instructions, examples, code, and supporting resources.

OpenAI currently documents skills for eligible Business, Enterprise, Healthcare, and Edu users, subject to workspace settings and product availability. Skills can be created with chat, an editor, or an uploaded skill package.

Open official ChatGPT skills guide ↗

Grok Bot skills

Save a successful task as reusable instructions across Bots.

Grok Bot skills can capture steps, decision rules, expected output, and safety boundaries. A separate routine defines when the workflow runs on a schedule or event.

Open official Grok Bot skills guide ↗

Use the idea anywhere

Even without a formal Skills button, document the process.

A clear skill document can become a team playbook, reusable prompt, internal SOP, agent instruction, or implementation specification.

Build one with private help →

Product information checked October 30, 2026. Eligibility, interfaces, sharing, and plan access can change; confirm with the official sources.

Turn successful work into a repeatable method

Bring one task. Leave with the workflow and the skill behind it.

The private AI Working Session is designed for exactly this: do the task together, capture what worked, use it again, then refine it in the included follow-up.

Keep going

Try it with a real task.

Put one idea into practice, then build on what works for you.