Have you written a long prompt, got the result you wanted, and then had to explain everything again the following week? The study-note format, caption style, or report structure is actually the same. Only the material and that day’s needs change. This is why learning to create AI skills is appealing: we save a clear way of working so it is easier to reuse.
In my earlier article on effective AI prompting, I discussed goals, context, audience, format, and constraints. Skills build on those principles. A prompt communicates the current need; a skill provides relatively stable rules and processes. They complement each other.
In technical terms, Agent Skills are packages of instructions and supporting resources for specific tasks. Their implementation differs between products, though. ChatGPT Work and Claude support skills, while Gemini offers Gems for saving guidance and reference materials. A Gem is not automatically the same kind of Agent Skills package. Here are eight tips for university students, creative professionals, business owners, community organizers, and office workers.
Create AI skills around one recurring need
Choose a task you perform frequently with reasonably clear standards for the result. Examples include summarizing lecture material, writing product descriptions, reviewing manuscripts, or preparing meeting follow-up actions. Avoid starting with a “help me with all my work” skill. Its scope is too broad and difficult to test.
Write down three things: the inputs usually available, the recurring process, and what counts as a finished result. A student might provide lecture material and want a summary and quiz. A shop owner provides product specifications and needs an accurate description, rather than exaggerated sales promises.
Example creation request: “Create a deskripsi-produk skill for a craft business. Its inputs are materials, dimensions, function, and care instructions. Its outputs are a short description, five benefit points supported by the specifications, and any information still needed. Don’t invent certifications or testimonials.”
A skill is worth creating when the same rules really are used often. For a one-off question, an ordinary prompt may be enough. Consistency also means maintaining process standards, rather than making every answer identical word for word.
Separate standing rules from each request’s needs
One easy mistake is putting too many temporary details into a skill. Event names, dates, products, and weekly targets usually change. If we save all of them as permanent rules, the next result may use outdated context.
Store the language, style, structure, constraints, and mandatory checks in the skill. Put the latest material and valid exceptions in the prompt. For creative work, standing rules might include a warm tone and a ban on inventing personal experiences. Supply the content’s topic and purpose when using the skill.
Skill rules: “Create three alternative captions in Indonesian, each no longer than 80 words. Use a friendly tone, avoid unsupported claims, and don’t add the author’s experiences.”
Daily prompt: “Use the caption-komunitas skill for this Sunday’s river cleanup. The audience is local residents; focus on asking people to bring gloves. Don’t mention the location until I confirm it.”
When you create AI skills, also explain how to handle conflicts. For example, a new request may change caption length, but must not override the ban on inventing facts. If the conflict is unclear, AI should ask which instruction takes priority.
Define the workflow and criteria you can check
When you create AI skills, explain what should happen to the material, rather than just assigning a role. “Be a great tutor” does not explain how to read the material, construct questions, or check the result before handing it over.
Use a short sequence: check the inputs, process the material, organize the output, then review it. Specify when AI should ask questions. If the material is unreadable, don’t produce a summary based only on its title. If the meeting date is missing, mark it as unknown.
Example study-skill workflow: “Read the supplied material. Explain the main concepts in simple language. Create five questions from that material. Show the questions first; provide answers and explanations after I try. If sources disagree, show the differences.”
Success criteria might include retaining key terms, basing questions on the material, and not revealing answers too early. For maintenance technicians, a similar workflow can organize inspection notes into findings and follow-up questions, without inventing diagnoses or replacing official safety procedures.
Starting with instructions alone is usually easier. Consider scripts or additional connections only when a real need calls for them.
Add focused examples, references, and constraints
Provide an example of a good result so AI understands our standards. One relevant example is often more useful than a string of words such as “professional, engaging, high-quality.” Explain which elements to imitate and which must not be copied.
A writer can provide a sample paragraph, an editor a terminology list, and a small business approved product information. Use material you are allowed to share. Don’t turn one example into an absolute rule when the next task has different needs.
Example editing rules: “Follow this example’s sentence rhythm, not its story. Preserve character names, viewpoint, figures, and direct quotations. Improve spelling and flow. Don’t change the ending or add the author’s personal experiences.”
According to the Agent Skills specification, a skill package contains a SKILL.md file and may include references, assets, or scripts. Here is an example of its basic structure:
--- name: catatan-belajar description: Summarize study material and create quizzes from user-supplied content; not for completing an active exam. --- Use only the supplied material. Produce a summary, key terms, and five questions. Withhold the answer key until the user has finished trying. If material is missing, ask; do not invent sources.
This is a starting example, not an installed package. Keep lengthy references separate and specify when to read them. You don’t need a long document to create AI skills; the instructions just need to be clear enough.
Adapt the setup for ChatGPT, Gemini, and Claude
To create AI skills across services, use the same process design, then adapt it to the service’s features. Don’t assume one ZIP file or command works immediately in every AI product. Feature availability, organizational permissions, and menus can change.
ChatGPT: use the skill creator in Work
OpenAI’s official skill-building guide explains using @skill-creator in ChatGPT Work. Open a supported environment, choose the skill creator through the @ mention menu, and describe the task, rules, examples, and output. Review the draft before saving and reusing it.
Example: “@skill-creator Create a ringkasan-rapat skill. From the user’s notes, organize decisions, responsible people, deadlines, and open issues. Don’t turn proposals into decisions. Use the following anonymized example as the format.”
Once available, select that skill for the next meeting and provide new notes. If the skill feature is not visible, saved instructions or a template offer a practical alternative; neither automatically becomes a skill package.
Gemini: use a Gem with saved instructions
For a similar approach, create a classic Gem in the Gemini web app through the Gems menu and the new Gem option. Enter a name, instructions, and reference materials in Knowledge if needed. Try a prompt in the preview, then save. Google’s Gems guide explains the steps; its Gem instruction tips emphasize tasks, context, roles, and format.
Instructions for the “Teman Belajar” Gem: “Help me understand the supplied material. Summarize concepts, explain terms, then ask five progressively harder questions. Don’t give the answers immediately. If the material is insufficient, ask.”
Open that Gem and send new material when studying. Instructions and examples can be adapted from a skill design, but using a Gem does not mean importing SKILL.md with all the package’s capabilities.
Claude: upload and enable a skill package
Prepare a folder containing SKILL.md and the necessary references, then package it as a ZIP. In Claude, open Customize → Skills, choose the skill creation or upload option, upload the package, and enable it. Follow the Claude Skills guide, including its code-execution environment and organizational permission requirements.
Example use: “Use the deskripsi-produk skill for this canvas bag. Material: canvas; dimensions: 30 × 35 cm; care: hand wash. The target readers are university students. Don’t say it is waterproof because that hasn’t been tested.”
Check that the skill is actually enabled. Simply typing its name in a chat does not mean the package has been installed or read.
Combine skills with prompts that stay specific
When we create AI skills, we can shorten our instructions, but the prompt still needs to state the current material and goal. Think of the skill as a basic recipe, while the prompt describes the available ingredients and necessary adjustments. Don’t expect AI to know that your needs have changed unless you tell it.
Here are some combinations to try once the relevant skill or Gem is available:
- University student, Gemini: open the Teman Belajar Gem, then write, “Use this chapter’s material. I already understand the definitions; focus on applications and five exercises, without the answer key for now.”
- Creative professional, Claude: “Use the penyunting-naskah skill on the following story. Its readers are teenagers. Make the dialogue more natural, but preserve the conflict and ending.”
- Office worker, ChatGPT Work: select the ringkasan-rapat skill, then write, “Process these notes. Distinguish decisions from proposals; flag tasks without an assigned person. Don’t guess deadlines.”
- Business owner: “Use the deskripsi-produk rules. Create a short catalog version from the following specifications. The price hasn’t been set; don’t include prices or discounts.”
If the result goes off track, determine whether the problem is in the skill or the prompt. Repeated formatting errors may require changes to the standing rules. An unclear topic for today’s task usually needs more context in the prompt.
Test normal cases, missing material, and different requests
After you create AI skills, don’t rely on them just because one result looks good. Test several inputs that represent real use. Include complete material, incomplete material, and requests that should fall outside the skill’s scope.
For a study skill, check that the quiz really comes from the material and that answers are withheld as instructed. For a shop, try a product with missing dimensions: does AI ask, or invent them? For meetings, provide notes containing proposals but no decisions: does it keep the two separate?
Example test prompt: “Use the deskripsi-produk skill. The only available information is that the material is cotton. Dimensions and care instructions are unknown. Identify the missing information before making product claims.”
Record simple assessments: factual accuracy, format compliance, preserved elements, and handling of uncertainty. Compare the result with what we expected. Revise one part of the rules, then test again, so we understand which change helped.
When we create AI skills, success means the results are easier to use and check. Don’t judge only the fluency or length of an answer. This kind of improvement process also relates to becoming an effective and reliable worker.
Maintain skills, protect data, and review results yourself
Give each skill a clear name, an update date, and a brief change log. When catalog formats, curricula, or working rules change, update the references. Keep test examples so a new version can be compared with the previous one. Archive unused skills to avoid a confusing set of choices.
Use AI wisely. Skills can store instructions, examples, and files for reuse; don’t put passwords, one-time passwords (OTPs), customer identities, children’s data, or confidential company data into shared packages. Use dummy examples and check organizational policies and service access.
If you use someone else’s skill, read its instructions and files first, particularly if it contains scripts or requests access to external services. The habit of securing digital accounts remains necessary. A skill does not authorize sending data, publishing content, or changing important documents without permission.
Example final rule: “Show a draft for review. Don’t send, upload, or publish the result. Flag unverified facts and sensitive information that needs reviewing.”
Check and recheck facts, sources, calculations, and tone. AI review helps, but people still need to read, decide, and add the finishing touches before using the result.
Conclusion: start with a small, clearly defined process
To create AI skills, start with recurring needs, focused rules, and examples of results you can evaluate. Adapt the setup to the service, then add a prompt explaining the current need. A short, clear skill is more useful than long instructions that contradict each other.
Let’s choose one routine activity: study notes, product descriptions, story editing, or meeting summaries. Set out the basic rules, test a few cases, and improve it gradually. The aim is to support a consistent process while preserving accuracy, confidentiality, and human judgment.
This post is also available in:
