AI Coworker Onboarding
Train an AI agent with context, constraints, examples, and iterative feedback.
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 94%
AI Coworker Onboarding treats a persistent agent as a new team member that must learn the company before receiving broad autonomy. Begin by importing information that can be gathered automatically, such as products, website copy, logos, colors, and brand guidelines. Then add the details automation cannot infer reliably: business goals, target metrics, disclaimers, visual rules, preferred actors, channel conventions, and examples of strong work. Review early outputs, explain corrections, and preserve recurring preferences so the agent does not repeat mistakes. A context or readiness score can expose missing knowledge and encourage deliberate improvement. The mechanism is straightforward: richer relevant context improves decisions, explicit constraints reduce violations, and iterative feedback turns a generic model into a company-specific collaborator.
Origin
Patrick Haiti compares configuring Supercale's agent to onboarding a human teammate and demonstrates a context score that rises as customers teach it company-specific rules. Extracted from Marketing Against The Grain.
Core principles
- 01Treat an agent like a new coworker rather than a context-free tool.
- 02Better context produces more useful and compliant outputs.
- 03Explicit constraints prevent repeated corrections.
- 04Examples communicate standards more clearly than abstract instructions alone.
- 05Agent readiness should improve through use and feedback.
How to run it
- 1
Import foundational context
Collect the company's products, website content, brand colors, logos, guidelines, and existing assets. Use automated ingestion where possible to avoid beginning from zero.
Pro tip Start with authoritative sources rather than scattered campaign drafts.
Watch out Automatically extracted context may be incomplete or outdated and still needs review.
- 2
Define goals and guardrails
Specify target outcomes, performance metrics, mandatory disclaimers, prohibited treatments, and budget or approval boundaries. Make ambiguous requirements explicit.
Pro tip Phrase constraints as observable rules the agent can check before completing work.
Watch out Do not grant publishing or spending autonomy before these boundaries are tested.
- 3
Teach preferences with examples
Provide representative successful work and explain preferred colors, logos, actors, formats, and channel conventions. Distinguish universal brand rules from campaign-specific choices.
Pro tip Pair each example with a brief explanation of why it is good or bad.
Watch out Conflicting examples can teach the agent inconsistent standards.
- 4
Review early work
Inspect initial outputs closely and correct both the visible error and the missing context that caused it. Ask the agent to explain important creative decisions when useful.
Pro tip Track repeated corrections; repetition indicates that a preference has not been preserved clearly.
Watch out Silently fixing outputs outside the agent prevents it from benefiting from the feedback.
- 5
Expand autonomy gradually
Increase task scope as the agent demonstrates reliable understanding of the company and its constraints. Continue updating its context as products, goals, and policies change.
Pro tip Use readiness scores as prompts for review, not as substitutes for testing real outputs.
Watch out A high context score does not eliminate the need for budget controls and human review of consequential actions.
In the wild
A regulated advertiser imports its brand materials, teaches the agent which ads require a disclaimer, specifies when logos may appear, and supplies acceptable creative examples. The team reviews initial work and adds missing rules whenever it detects a recurring error.
→ The agent produces more consistent campaign assets with fewer compliance and branding corrections.
A brand identifies approved characters from an avatar library, explains which audiences and campaigns each character suits, and saves those choices in the agent's context. Early videos are reviewed before the agent receives permission to create larger batches.
→ Subsequent campaigns use a more consistent cast and recognizable brand personality.
Common mistakes
Skipping onboarding
A capable model still lacks private company knowledge. Asking it to operate immediately produces generic work and avoidable errors.
Providing context without priorities
A large context dump can contain contradictions or irrelevant material. Identify authoritative rules and explain which instructions take precedence.
Equating confidence with permission
A readiness score indicates supplied context, not guaranteed judgment. Consequential publishing and spending actions still require tested controls.
Is it for you?
Best for
It is best for teams adopting persistent AI agents that repeatedly create, analyze, or distribute marketing work.
Not ideal for
It is not ideal for isolated, low-stakes tasks where persistent company context offers little benefit.
From the transcript
“you can actually teach your system what to So, we have customers that are using the system and so they're putting in like really nuanced…”
“they're literally doing like an onboarding type of, right? Like you would onboard a new person to your team.”
“The way that I think about it is like you would never have someone joined your team as a human without an onboarding, right?”
From the episode
We Tested an AI Agent That Builds 1000 Ads in 10 Minutes