How to Train Claude AI: Teach a Task or Use the API

Can you train Claude AI?

“Training Claude” can mean changing the model itself or giving Claude instructions to follow when you use it. Prompts, Skills, and API instructions guide Claude at use time; they are not the same as changing the model’s weights. Anthropic’s glossary says the Claude API does not currently offer fine-tuning and advises contacting Anthropic to explore the option. Check its latest guidance before making a decision. (Anthropic glossary)

For a repeatable task in Claude, you can create a Skill. For an application, you can send instructions with an API request. These approaches guide Claude for a task; they are not fine-tuning. (Claude Skills documentation)

Choose how to guide Claude

What you need	Use this	Why
Help with a task once	A detailed prompt	Explain the goal, context, and desired output without creating reusable setup.
A repeatable workflow in Claude	A Skill	Package task instructions, standards, and examples for relevant work.
General preferences across chats	Custom instructions	Use these for broad preferences rather than a detailed task procedure.
Work with context that builds over time	A Project	Keep project documents and conversation context together; add a Skill if you also need a repeatable procedure.
Put Claude in your own application	The Messages API	Send task instructions and user input from your application.

Anthropic recommends direct prompting for one-off or exploratory tasks, Skills for repeatable procedures, and Projects for work that needs accumulated context. (Claude Academy)

Teach Claude a recurring task with a Skill

A custom Skill is a folder with a required SKILL.md file and optional reference files or scripts. Claude can load it when a request matches the Skill’s description. (Custom Skills documentation)

1. Check access first. Skills require Code execution to be enabled. For Free, Pro, Max, and Ultra, check Settings > Capabilities; in the updated Settings layout, file creation may be called Create and edit files. For Team and Enterprise, an Owner manages the relevant settings under Organization settings > Plugins & skills, on the Policy tab. (Claude Help Center)
2. Choose a task you repeat. For example, turning rough notes into a weekly team update.
3. Describe what good looks like. Specify the audience, format, tone, rules, and an example you consider successful. Ask Claude to draft a Skill and review its instructions.
4. Package the Skill. Put SKILL.md inside a folder named for the Skill. Add any supporting files its instructions refer to, then package the folder as a ZIP. Anthropic’s documentation describes the Skill folder structure; its Help Center’s upload instructions specify a ZIP containing SKILL.md. (Custom Skills documentation; organization upload instructions)
5. Upload and turn it on. Open Customize > Skills and upload your personal Skill. When it appears in your Skills list, select Turn on on its row, or open it and use the switch on its page. (Claude Help Center; Skills overview)
6. Test in a fresh conversation. Ask for the task normally. The Claude Academy tutorial suggests checking Claude’s thinking for “Reading [your skill’s name].” If it did not load, improve the Skill’s name or description, or mention it by name. Compare the result with your example and revise instructions where needed. (Claude Academy)

Example prompt for a weekly-update Skill:

Create a Skill for my weekly team update. Use three sections: wins, blockers, and priorities, with no more than three bullets in each. Write in a direct, factual tone for a manager to skim. Use this example as the quality bar: [paste an update you liked].

This follows Anthropic Academy’s suggested pattern: name the format, set the tone and audience, and provide a good example. It is a starting point, not a guarantee that every output will be correct. (Claude Academy)

Guide Claude through the production API

For an application, send the task instructions and the specific task in a user message. The Messages API accepts a list of messages with a role and content; its documentation shows a single user message as a valid request. (Messages API reference)

You need a Claude Console account and API credentials. Install Anthropic’s Python SDK, set ANTHROPIC_API_KEY in your environment, and keep the key out of source code and version control. The SDK’s documented Python pattern uses client.messages.create; Anthropic’s SDK documentation provides the model identifier used below. (API overview; Python SDK)

pip install anthropic
export ANTHROPIC_API_KEY="your-api-key"

Save this as guide_claude.py:

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=500,
    messages=[
        {
            "role": "user",
            "content": (
                "Turn these rough meeting notes into a concise weekly update for a manager. "
                "Use exactly three headings: Wins, Blockers, and Priorities. "
                "Include no more than three bullets under each heading. Be direct and factual. "
                "If the notes do not support a detail, do not invent it.\n\n"
                "Notes: finished onboarding; waiting on analytics access; "
                "draft launch checklist next week."
            ),
        }
    ],
)

for block in message.content:
    if block.type == "text":
        print(block.text)

Run it with:

python guide_claude.py

This example puts the instructions and task together in the user message; it does not train the model. Anthropic’s Messages API guide also documents a top-level system field for instructions that apply from the start, if you choose to use that structure instead. (Messages API guide)

Test and improve the result

Before reusing a prompt or Skill, define what a good result means for the task—for example, correct headings, no invented facts, and no more than three bullets per section. Try several representative inputs, including an incomplete or unusual one. Note which criteria fail, revise the instructions, and test again. For a Skill, also confirm it loaded and compare its output with your example. This is a practical iteration loop, not a guarantee of a particular result.

Sources

Glossary — Claude Platform Docs

Teach Claude your way of working using skills — Claude Academy

Create custom skills — Claude.ai Documentation

Use skills in Claude — Claude Help Center

Provision and manage skills for your organization — Claude Help Center

Skills overview — Claude.ai Documentation

Create a Message — Claude API Reference

API overview — Claude Platform Docs

Python SDK — Claude Platform Docs

Using the Messages API — Claude Platform Docs