To train employees to use Claude, first say which Claude account they may use for work and what must stay out of it. Then have each person practice on their own role’s work: which jobs to hand to Claude, how to brief it, what to show it and how to check its answer. Once a job repeats, teach them to write it down as a skill, a SKILL.md file Claude can use whenever a request matches.
Set the rules for Claude at work
Before anyone practices, write down:
- Which account. A company Team or Enterprise plan, an individual plan, or neither for certain work. On company plans, owners control settings such as connectors and skills.
- What stays out. Passwords, client personal data, anything marked confidential. Give examples from your own business.
- Which connectors are allowed. Claude can connect to other tools. Say which ones people may connect, and that they should read the permissions a connector asks for.
- What gets checked, and by whom, before Claude-assisted work leaves the team.
Practice what to hand Claude, role by role
Judgment comes first: which jobs go to Claude, which need checking first, and which need a different tool. Practice it on each role’s work, not a generic example:
- A project manager might hand Claude a summary of a long email thread, and check every date before it goes on the schedule.
- An accountant might hand it the wording of a variance note, and keep the numbers themselves in the spreadsheet.
- An account manager might hand it a first draft of a renewal email, and check the terms against the contract.
Practice briefing Claude
A brief says what you want back, who it is for, the limits and what counts as done. Laura, an operations lead, used to ask Claude to “tidy up these notes.” Now she asks for “a half-page summary for the site team, decisions first, open questions last, nothing that isn’t in the notes.” She gets fewer rounds of edits. See how to write a good prompt.
Practice giving Claude the right context
Claude works from what it can see: what is in the chat, the files you add, and what its connectors reach. Practice choosing the one document it needs, leaving out what your rules keep out, and knowing that a new chat may not remember the last one. Practice, too, knowing what a connector can reach before you rely on it. See how to give AI the right context.
Practice checking Claude’s work
Claude can state something false as if it were true. Have people practice on drafts with a mistake in them, against the facts the draft was based on, and say what doesn’t hold up. The habit to build is reading before sending. See when to trust what AI tells you.
Build skills for the jobs people repeat
This is where Claude training goes further. A Claude skill is a folder with a SKILL.md file: a name, a description of when to use it, then plain instructions. Claude keeps only the name and description in view, and reads the rest when a request matches. So a job explained once can be reused without explaining it again.
Teach people to build one in this order:
- Pick a job they have explained to Claude more than twice.
- Write what starts it, what it works from, the steps and what it hands back.
- Write the description from the words they actually use when they ask for that job.
- Say what done looks like, and what the skill must never do, such as send anything on its own.
- Run it on a few real cases, including a messy one, and fix the instruction behind each mistake.
To add a skill in Claude, zip its folder, then go to Customize, Skills, and upload it. Anthropic’s help center says the feature needs code execution turned on; on Team plans skills are on by default, and on Enterprise an owner turns them on in organization settings. A skill still needs checking: it gives Claude the same instructions, not a guarantee of the same right answer.
Setting up training inside Claude
Training can run as an app inside Claude, added as a custom connector. On an individual plan, go to Customize, then Connectors, and add a custom connector with the service’s address. On Team and Enterprise plans, an owner adds it in organization settings, and each person then connects it from their own account. Anthropic’s help center has the current rules for each plan, and how AI training inside ChatGPT and Claude works covers what such an app can see.
Follow up after the training
- Look at the calls people miss, and bring a shared one to the next team meeting.
- Share good skills. When someone builds a skill that works, have them show the team, and have others read every file before adding it.
- Measure decisions, not logins. See how to measure how well people use AI.