#022 Agent Skills: How to Turn Repeated AI Work into a Reusable System

What agent skills are, where they are useful, and how to build one for ChatGPT or Claude.

People who use AI regularly often develop methods that work. Few preserve those methods in a form they can reuse.

A product manager keeps a long prompt for reviewing proposals. A marketer pastes the same brand guidance into every campaign request. A founder explains the business, customer and market again before asking for a strategy memo. Their methods work, but remain buried in personal chat histories.

Agent skills provide somewhere to store that method.

A skill packages the instructions, knowledge, examples and checks behind a recurring task. When that task returns, the AI can load the package. The user provides the facts for the current job without rebuilding the method from scratch.

This addresses more than wasted time. Many teams store their best processes across old documents, message threads, personal habits and the memories of a few experienced employees. Skills can make part of that knowledge reusable.

The same mechanism can preserve bad assumptions, expose sensitive information or impose a rigid process on work that requires judgment. The method inside the skill determines its quality.

What this guide covers

• What an agent skill is

• How a skill differs from a prompt, project, custom assistant and agent

• How skills are selected and loaded

• Which workflows are worth packaging

• How to design a useful skill from one real task

• How to create and use it in ChatGPT and Claude

• How to test, secure and maintain it

• Why the idea is relevant to African teams, and where its limits remain

What is an agent skill?

An agent skill is a folder that teaches an AI system how to handle a specific type of work.

Every skill folder needs a SKILL.md file containing its name, description and working instructions. The folder may also hold scripts, reference documents, templates and other assets. Several agent products support this structure through the open Agent Skills specification.

A basic folder might look like this:

kenyan-coffee-campaign/
├── SKILL.md
├── references/
│   ├── brand-guide.md
│   ├── approved-product-claims.md
│   └── customer-segments.md
└── assets/
    └── campaign-brief-template.md

The SKILL.md file explains the task and method. The other files supply information the AI may need while doing the work.

A skill keeps the operating method in SKILL.md and adds supporting material only when the workflow needs it.

A skill works like an operating playbook. It can define:

• When the workflow should start

• Which inputs are required

• Which steps the AI should follow

• How it should make decisions

• What the final output should contain

• What it must check before finishing

The model supplies broad language, reasoning and generation capabilities. The skill supplies the method for this particular job.

How a skill works

Skills use a pattern called progressive disclosure. The AI initially sees only the name and short description of each available skill. It opens the full instructions when a request appears relevant, keeping unrelated guidance out of the conversation.

A skills-compatible agent follows three stages:

1. Discovery: It sees the name and description of the available skills.

2. Activation: When a request matches a skill’s description, it loads the full SKILL.md instructions.

3. Execution: It follows those instructions and opens supporting files or runs scripts when the task requires them.

Progressive disclosure keeps the full instructions out of the conversation until the request makes them relevant.

The description acts as a routing instruction. It tells the AI when the skill belongs in the work. An overly broad description can activate the skill for unrelated tasks. A vague description can stop it from appearing when needed. The Agent Skills guidance on descriptions recommends testing requests that should activate the skill and requests that should not.

In ChatGPT and Codex, users can select a skill directly, or the product can select it when a request matches its description. ChatGPT uses @ mentions for explicit selection, while Codex uses $ mentions. OpenAI’s current documentation also describes a built-in skill creator for turning a workflow into a package.

Claude reviews the available skills, loads relevant instructions and applies them to the task. Its upload and management flow differs from ChatGPT’s. The underlying principle remains the same: load the operating knowledge required for the work. Anthropic’s skills overview explains this loading model.

How a skill differs from a prompt

A prompt tells the AI what to do in the current conversation. A skill stores the method for handling a recurring type of work across conversations.

An example makes the distinction clearer.

Prompt:

Write three Instagram posts for our new coffee blend. Use a warm tone and mention that the beans come from Nyeri.

That request may be enough for a one-off draft. The user supplies the task, context and desired output together.

Skill:

The skill already contains the brand voice, approved product claims, audience segments, campaign process, channel rules and final review checklist. The new request only needs to provide the campaign brief and current product facts.

Skills sit alongside several related AI features:

ToolMain jobBest used whenPromptInstruct the AI in the current conversationThe request is quick, one-off or exploratoryProject or workspaceKeep related conversations and files togetherThe user needs continuing context around a topic or body of workCustom assistantConfigure an AI experience for a broad role or audienceThe same assistant will handle several related tasksSkillPackage the method for one recurring workflowThe steps, standards and output need to remain consistentAgentPursue a goal through reasoning, tools and multiple actionsThe work requires decisions and execution across several steps or systems

The categories overlap. An agent may use several skills; a project may store the files they need; and a custom assistant may have access to them. The useful question is which tool is sufficient for the job.

Which work deserves a skill?

Repetition is only the starting point. “Answer customer emails” is too broad to encode well. “Draft a first response to delivery-delay complaints using the approved resolution policy” has a workable boundary.

Four conditions make a task a strong candidate:

1. It repeats. The task returns often enough to justify preserving the method. Examples include a weekly campaign brief, monthly portfolio review, recurring research report or standard document check.

2. It needs stable context. The AI repeatedly needs the same brand rules, product definitions, approval criteria, market assumptions or output format.

3. The team can define good work. If experienced colleagues cannot agree on the process or quality standard, a skill will encode the disagreement rather than solve it.

4. Inconsistency has a cost. That cost could be rework, a weak decision, an unsupported marketing claim, a missed compliance check or weeks spent teaching a new employee an undocumented process.

Skills are less useful for novel strategy, sensitive judgment calls and tasks whose process changes every time. Missing data, unclear ownership and broken processes need to be fixed before a skill can help.

Worked example: a campaign skill for a Kenyan coffee brand

Consider a Kenyan coffee brand that launches a new campaign each month. The product, offer and campaign objective change, but much of the working method stays stable.

The team always needs to understand the product, identify the relevant customer segment, connect the product to a credible customer need, develop one campaign idea and adapt the message for Instagram, email and the website. Every claim must come from approved product information. The copy should reflect the brand’s voice and Kenyan context without falling into stereotypes or forcing slang.

The task is specific: Create a channel-ready marketing campaign from an approved coffee product brief.

“Help with marketing” is too broad. It could refer to pricing, market research, media buying, event planning or customer support. A narrower skill is easier to activate and test.

The campaign facts change each month. The brand rules, process and checks remain inside the skill.

Here is a simplified SKILL.md:

---
name: kenyan-coffee-campaign
description: Use this skill when creating a marketing campaign from an approved product brief for the Kenyan coffee brand. It applies to campaign concepts and copy for Instagram, email and website launches. Do not use it for pricing, media buying, customer-support responses or unapproved product claims.
---

# Campaign workflow

## Required inputs

Ask for the following if they are missing:

- Approved product brief
- Campaign objective
- Target customer segment
- Offer and campaign dates
- Required channels

## Process

1. Read the product brief and approved claims.
2. Identify the customer need relevant to the selected segment.
3. Develop one campaign idea and explain why it fits.
4. Write the core message before adapting it by channel.
5. Produce Instagram, email and website copy.
6. Check every factual claim against the approved source.

## Standards

- Use clear, warm language.
- Reflect Kenyan context only where it is relevant and supported.
- Do not invent origin, quality, health or sustainability claims.
- Do not force slang or treat Kenyan customers as one audience.
- Keep the same campaign idea across all channels.

## Output

Return:

1. Campaign objective
2. Target segment and customer need
3. Campaign idea
4. Core message
5. Channel copy
6. Claim and compliance check
7. Missing information or unresolved questions

The supporting references/ folder could hold the brand guide, approved claims and customer-segment definitions. Keep the main file focused. Put detailed material in reference files that the AI can open when needed.

The skill still needs the approved brief, current offer, dates and chosen audience. It preserves the method, but accurate input remains necessary.

Designing your first skill

Start with work that has already produced a good result. A proven campaign process gives you something concrete to encode; a universal marketing method does not.

Write down six decisions before touching the file:

1. Task: What single job should the skill complete?

2. Trigger: Which user requests should and should not activate it?

3. Inputs: Which facts, files or decisions must the user provide?

4. Process: Which steps must happen, and in what order?

5. Standards: What must the AI include, avoid, verify or escalate?

6. Output: What should the completed work look like?

Match the instructions to the risk. A brainstorming skill can leave room for exploration. A skill checking financial claims or regulated communication needs approved sources, clear approvals and explicit stopping conditions.

Examples help when a rule is hard to explain. One approved campaign and one rejected campaign may show the difference between warm brand language and inflated product claims better than another page of adjectives.

Add scripts only when the task needs a fixed calculation, file conversion, validation check or tool connection. Instructions are enough for many first skills. Both OpenAI and Anthropic recommend starting with a focused, instruction-led version.

Setting it up in ChatGPT and Claude

ChatGPT and Claude support the same broad skill structure, but their setup flows differ.

ChatGPT Work and Codex

In ChatGPT Work:

1. Invoke @skill-creator.

2. Describe the workflow you want to preserve.

3. Answer questions about the task, trigger and required files or scripts.

4. Review the generated package.

5. Test it with realistic requests.

You can create the folder and SKILL.md manually. Either way, review the package as a first draft.

Once installed, ChatGPT can select the skill when a request matches its description, or the user can choose it explicitly with an @ mention. Codex uses the same broad skill structure, with $skill-creator for creation and $ mentions for direct selection. Access can depend on the user’s plan, platform, region and workspace settings, so the visible controls may differ.

Claude

For Claude:

1. Create the skill folder.

2. Add the SKILL.md and supporting files.

3. Package the folder as a ZIP file. Anthropic’s current instructions require the ZIP to contain the skill folder at its root.

4. Open Customize → Skills → + Create skill → Upload a skill.

5. Upload and enable the skill.

6. Test it with realistic requests.

After upload, test prompts that should trigger the skill and confirm Claude loaded it. Team and Enterprise administrators can also provision approved skills across an organization, subject to the relevant workspace controls.

Menus will change. The durable parts are the folder, description, instructions, resources and tests.

Test the skill like a product

A single successful prompt proves little. The base model may have handled the task on its own, or the prompt may have supplied the missing context by accident.

Test two separate things: activation and output quality.

Test whether the skill enters the right work before judging whether it improves the result.

For activation, prepare requests that should use the skill and requests that should not. Change the wording and level of detail. For the coffee example:

• Clear match: “Create the launch campaign for our new Nyeri blend.”

• Less explicit match: “Help us market this product,” with a product brief attached.

• Non-match: “Calculate delivery costs for our wholesale orders.”

For output quality, define the expected result before testing. Check whether the skill:

• Asked for missing inputs

• Followed every required step

• Used the correct reference files

• Kept the campaign idea consistent across channels

• Flagged unsupported claims

Then compare the same tasks with and without the skill. If the skill does not improve consistency, reduce rework or prevent a known failure, adding it has created maintenance without creating value.

The Agent Skills evaluation guidance recommends using realistic prompts, explicit success criteria and optional input files. Teams with higher-risk workflows should retain the test set and rerun it whenever the skill changes.

Where skills fail

Five failures appear repeatedly:

1. The scope is too broad. A single “marketing skill” becomes a dumping ground for strategy, research, content, media planning, reporting and customer engagement. The instructions conflict, and the AI struggles to identify which rules matter. Several focused skills usually work better.

2. The process is unclear. If the team has not agreed on the audience, approval process or quality standard, the skill cannot create that agreement. Polished output may hide the disagreement.

3. The knowledge becomes stale. Prices, policies, product facts, regulations and brand guidance change. Every skill needs an owner, a review date and an update process. Version control matters once several people depend on it.

4. Consistency creates false confidence. A skill can produce consistent work that is still wrong. Human review remains necessary where claims, customer harm, money, legal duties or reputation are involved.

5. The skill creates a security risk. Skills can contain scripts, packages, references and instructions that access external systems. Anthropic identifies prompt injection and data exfiltration as key risks. It advises users to install skills only from trusted sources, review bundled files and avoid hard-coded secrets. Its Enterprise scanning feature can flag some malicious third-party skills, but a passing scan does not guarantee safety. Inspect what a skill can read, run and send before enabling it.

Why this matters for African teams

Skills become useful when formal documentation is thin, experienced staff carry crucial process knowledge and global AI systems lack local context. Those conditions exist in many organizations, including teams working across African markets.

A Kenyan campaign team may need approved language for county-level activations, local product availability and customer segments that do not map neatly onto a global template. A Nigerian logistics team may have exception rules shaped by its own routes, payments and merchant operations. A South African service team may need a different language, policy and regulatory base. These are separate operating contexts, not one “African” instruction set.

Keeping guidance close to the task can help a smaller team share one tested method without building a large software system. That conclusion follows from how skills package procedural knowledge. It does not prove widespread adoption by African firms; reliable public data on their use across the continent remains limited.

Skills cannot repair missing source data, unreliable access, weak governance or a process built around undocumented exceptions. They can only work with what a team can state, supply and maintain.

Global AI products often arrive with examples shaped by different organizations, markets and assumptions. Product Pulse Africa tests where a capability fits, adds the local judgment it lacks and states what still needs human ownership.

What to do next

Choose one recurring task that caused rework during the past month. Find one example that your team considers good. Write down the inputs, steps, standards and final format that produced it. Add the situations where the method should not be used.

Build the smallest instruction-only skill that can reproduce that method. Test it on:

• Three normal requests

• Two awkward requests

• Two unrelated requests

Keep it only if it improves the result against a standard you defined beforehand.

Then share your experience. We want to hear:

• What you built the skill to do

• Where it worked well

• Where it failed or needed more guidance

• What you changed after testing it

• Whether it improved the work enough to keep using

Your feedback will help us compare how agent skills perform across real product, research, marketing and operational work in different African markets.