Write the complete specification for an AI image generator to create the hero visual for the Steadfast landing page. Include every detail the model needs — composition, color palette, mood, subjects, art style, aspect ratio, what to explicitly include, and what to explicitly avoid. The AI cannot read your mind; it will interpret your words literally.

Think about what happens when you leave something unspecified — the AI fills in the blank with its best guess, which may not match your vision.

Before writing any more prompts, answer: What do you believe makes a fintech product genuinely trustworthy — not just trustworthy-looking? What specific values should this campaign communicate? How would you encode those values into your AI instructions so the outputs reflect them consistently?

Your beliefs and values should drive your AI specifications, not the other way around.

Map the full deliverable set across modalities. For each piece (hero image, landing page copy, explainer video script, social posts, email drip), specify: which AI tool or model you would use, the output format and dimensions, the specific acceptance criteria for "done," and how you ensure visual and thematic consistency across all five deliverables.

Consistency across modalities is where most AI-assisted campaigns fall apart.

Identify every quality issue — factual problems, reasoning flaws, unsupported claims, tone issues, and anything that would erode reader trust. For each issue, explain what type of error it is and why it matters.

Look beyond grammar. The most dangerous errors are the ones that sound authoritative but are wrong or misleading.

Design an automated quality harness for this pipeline. Define at least 4 specific checks you would encode, describe how each one works technically, what threshold or signal triggers a flag, and what happens when a check fails (block, flag for review, auto-correct, etc.). Be precise enough that an engineer could implement your design.

Good quality harnesses catch categories of errors, not just individual instances.

This report quality has been degrading gradually over two weeks — not a sudden break, but a slow decline in accuracy and tone. What are the three most likely root causes of gradual degradation in a multi-agent content pipeline? For each, describe the diagnostic indicators you would look for and the specific corrective loop you would implement to prevent recurrence.

Gradual degradation has different root causes than sudden failure. Think about what changes slowly in an AI system.

Decompose this system into distinct agents. For each agent, define: its name and specific responsibility, what it receives as input, what it produces as output, how it decides to hand off to another agent, and what constraints or guardrails limit its behavior. Explain why you chose this decomposition over alternatives.

Think about separation of concerns. An agent that does too many things fails in unpredictable ways.

Design the context architecture. The documentation database is 50,000 pages. Each conversation may reference previous interactions. Token windows are finite. How does the right information reach the right agent at the right time? Address: retrieval strategy, context window management, conversation memory design, and how you prevent agents from being overwhelmed with irrelevant context.

An agent with too much context is as broken as one with too little. The architecture is in what you exclude.

This system accesses customer billing data and personal information under SOC 2 and GDPR. Design the trust and security architecture: What data can each agent access (and what can it NOT access)? How do you prevent prompt injection attacks from customers? How do you audit what agents do? What happens if an agent hallucinates a billing action or leaks PII?

Security in agentic systems isn't just authentication — it's about what a probabilistic system is allowed to do with certainty-requiring data.

Where in this system must humans remain in the loop? Identify at least 3 specific decision points that should never be fully automated. For each, describe how you design the handoff so the human agent has full context to act immediately — not spend 10 minutes catching up on what happened before them.

The hardest part of human-in-the-loop design isn't identifying where humans go — it's ensuring they can be effective when they arrive.

Calculate: (1) The daily token cost for Tier 1 and Tier 2 separately. (2) The total monthly cost of the entire AI system — including token costs, infrastructure, and remaining human agents for Tier 3 escalation. (3) The monthly savings compared to the current all-human team. (4) The break-even point: at what daily conversation volume does the AI system cost the same as the current team? Show your work.

Precision matters. In enterprise proposals, a wrong decimal point can kill a deal or greenlight a losing investment.

Define three specific, measurable goals for this AI system's first 90 days in production. For each goal, specify: the exact metric, the target value, how you'll measure it, the baseline (where it starts), and what specific action you'll take if the goal isn't met by day 90.

Good goals are specific enough to know when you've hit them and clear enough to know what to do when you haven't.

Beyond raw cost savings, identify the three biggest risks to this ROI projection. For each risk: explain what could go wrong, quantify or estimate the financial impact if it happens, and describe a specific safeguard you would build into the system design — before launch, not after — to protect the business case.

The best business cases don't just show the upside — they show you've already accounted for what could go wrong.

Look back at your responses across this entire assessment. Which response are you most confident in, and why? Which felt weakest, and what specific skill gap does that reveal? Be honest — self-awareness scores higher than self-promotion.

The ability to accurately assess your own capabilities is itself a core Human Velocity skill.

Write a specification for an AI agent that would help you practice the skill you identified as weakest. Define: what the agent does, what inputs it takes, what outputs it produces, how it gives you feedback, what guardrails keep it safe, and what success looks like after 30 days of use. Be precise enough that an engineer could build it from your description alone.

This tests two things at once: can you write a precise specification, and do you know what you need to learn?