Teach the AI What Good Looks Like
Onboard AI with explicit standards, examples, and task documentation.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 99%
This framework manages an AI assistant like a capable new employee who still needs role-specific onboarding. The manager first defines the task, desired outcome, and constraints, then documents the operating process and supplies examples of successful work. Crucially, the documentation should make implicit expertise explicit: what decisions matter, what quality looks like, and why an example succeeds. The AI is then tested against that standard, with failures used to improve the onboarding material rather than immediately blaming the model. Because the same clarity benefits human employees, the process can expose organizational ambiguity and clarify which responsibilities belong to people versus machines. The result is a reusable knowledge layer that improves consistency and enables later automation.
Origin
Extracted from Marketing Against The Grain, illustrated by Klarna improving AI customer support after rewriting inadequate onboarding documentation.
Core principles
- 01AI output reflects the quality of its management.
- 02Clear documentation benefits humans and AI together.
- 03Examples make quality standards observable.
- 04Domain experts gain leverage by teaching their craft.
- 05Operational clarity should precede automation.
How to run it
- 1
Define the Assignment
State the role, task, audience, constraints, and desired outcome. Make the scope narrow enough to evaluate.
Pro tip Describe the assignment as if onboarding a new employee tomorrow.
Watch out A vague objective such as creating a good strategy is not an actionable brief.
- 2
Capture the Process
Document how an expert completes the task, including inputs, decisions, handoffs, and exceptions. Resolve contradictions in the current process.
Pro tip Writing for AI can reveal ambiguity that humans previously worked around informally.
Watch out Do not encode an undocumented process from memory without checking how the work is actually done.
- 3
Show Successful Examples
Provide representative examples of completed work and identify the properties that make them good. Include enough variation to prevent imitation of only one surface style.
Pro tip Use examples from the real operating environment whenever possible.
Watch out Examples without explanations can teach superficial patterns rather than the underlying standard.
- 4
Test the Onboarding
Give the AI realistic assignments and compare its output against the documented criteria. Record where interpretation or execution fails.
Pro tip Have a domain expert evaluate substance rather than polish alone.
Watch out Do not interpret every failure as a model limitation before checking the brief.
- 5
Rewrite for Clarity
Update the instructions, examples, and criteria based on observed failures. Retest until performance is consistent enough for the intended use.
Pro tip Use the revised material to improve human onboarding as well.
Watch out Scaling before consistency is established multiplies errors.
- 6
Assign Human and AI Roles
Use the clarified workflow to decide what AI can perform, what humans must review, and what remains exclusively human. Document those boundaries.
Watch out Do not remove human judgment from high-impact decisions merely because the task is documented.
In the wild
Klarna's early customer-support AI integration underperformed even though the models and implementation were capable. The company rewrote the onboarding documentation used for human support representatives, giving the AI a clearer model of the task and expected quality.
→ The AI assistants improved, and the deployed support capability was described as equivalent to roughly 700 support agents.
Instead of typing only a request for a good marketing strategy, a marketing leader documents what a good strategy contains, provides representative examples, explains the market context, and defines how proposals will be judged.
→ The AI can produce a strategy grounded in the leader's actual domain expertise rather than a generic answer.
Common mistakes
Blaming the Model First
Teams may attribute poor output to AI limitations when the real failure is incomplete onboarding or an undefined quality bar.
Keeping Expertise Implicit
An expert who cannot explain the craft gives the assistant no reliable mechanism to reproduce it.
Documenting Only the Happy Path
Instructions that omit exceptions and judgment calls can fail as soon as the AI encounters realistic variation.
Is it for you?
Best for
Organizations delegating repeatable knowledge work to employees, assistants, or AI agents.
Not ideal for
Exploratory tasks where no reliable process, evaluation criteria, or examples yet exist.
From the transcript
“How you be successful with AI is you have to teach it what good looks like.”
“Once they rewrote the onboarding docs for the human, the AI agents all got better because now they understood how to do that task and…”
“usually your results with AI are reflection of you as a manager of how you are managing that AI assistant.”
From the episode
From Beginner to AI Expert in 30 Minutes: The 5-Step Framework You Need