Onboard AI Like an Employee
Teach an AI role through onboarding documents, projects, and reusable skills.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 97%
Treat a custom AI assistant as a new employee who needs structured onboarding rather than isolated commands. Begin with a role or job description, identify the skills and standards needed for that work, and create onboarding documents that explain procedures, examples, boundaries, and expected outputs. Install those materials in a persistent environment such as a custom GPT or Claude Project, then add modular skills that can be invoked when relevant. Test the assistant on representative assignments and revise the onboarding assets when systematic failures appear. The mechanism converts tacit expectations into durable context, increasing consistency across repeated tasks.
Origin
Kieran Flanagan described creating onboarding documents for custom GPTs, while Kristen Fraccia described Claude Projects that pull from reusable marketing skills.
Core principles
- 01AI performance improves when role knowledge is made explicit.
- 02Reusable onboarding beats repeated ad hoc prompting.
- 03Skills should activate when relevant to the current task.
- 04Job descriptions can seed role-specific capability sets.
How to run it
- 1
Define the role
Write the AI’s responsibilities, users, recurring tasks, boundaries, and success criteria.
Pro tip Start from a real job description when appropriate.
Watch out Do not assign conflicting responsibilities to one assistant.
- 2
Build the onboarding document
Explain domain concepts, workflows, quality standards, tools, and representative examples.
Pro tip Ask a research model to help structure the curriculum.
Watch out Review generated training material for fabricated guidance.
- 3
Create modular skills
Separate recurring procedures into focused instruction sets that can be applied when relevant.
Pro tip Give each skill a clear trigger and output contract.
Watch out Overlapping skills can produce contradictory behavior.
- 4
Install persistent context
Load the onboarding materials and skills into a custom GPT, project, or comparable persistent assistant.
Pro tip Version the assets outside the AI interface.
Watch out Confirm the platform’s privacy and retention settings.
- 5
Evaluate and retrain
Run realistic tasks, identify repeated errors, and revise the relevant onboarding document or skill.
Pro tip Use fixed test cases to detect regressions.
Watch out Do not patch every bad output with another global rule.
In the wild
A marketing leader creates a Claude skill that identifies AI-writing patterns, then makes it available inside multiple Claude Projects. When writing tasks arise, the projects can apply the editor’s standards without receiving the full instructions again.
→ Marketing drafts become more consistently aligned with the team’s preferred human style.
Common mistakes
Using only a broad persona
A role label without procedures, examples, and standards does not provide enough operational context.
Letting skills contradict
Multiple instruction modules need clear scopes and priorities or they can produce unstable behavior.
Is it for you?
Best for
It is best for recurring knowledge-work roles with stable standards, examples, and procedures.
Not ideal for
It is not ideal for one-off requests or roles whose source material cannot safely be placed in the selected AI system.
From the transcript
“if you kind of build onboard and docs for custom GPTs, you can onboard them to skills in the same way you would with an…”
“all of my Claude projects now pull from Claude's skills when they're relevant.”
“you could upload a job description and you can have the app give you a bunch of skills based upon that job description that you…”
From the episode
The AI Workflow That Lets 50 People Do the Work of 500 ($2B Founder Reveals)