The AI Fluency framework, developed by Anthropic as part of its AI Fluency course, defines effective human-AI collaboration through four core competencies — the 4Ds — paired with three interaction modes that describe who is driving at any given moment. It is a vocabulary for talking about how well a person or team works with AI, complementing organizational frameworks like the dora-ai-capabilities-model that describe conditions at the system level.
The 4Ds
- Delegation — knowing what to hand to AI and what to keep. Good delegation matches the task to the tool (and to the interaction mode), scopes the request crisply, and supplies the context the AI needs to succeed
- Description — articulating intent, context, and constraints clearly enough for the AI to act on them. The quality of the description sets the ceiling on the quality of the output
- Discernment — judging AI output critically: is it correct, is it appropriate, does it actually solve the problem. Discernment is what prevents plausible-but-wrong output from flowing downstream unreviewed
- Diligence — the care around how AI is used: verifying claims, maintaining traceability, respecting privacy and IP, owning the final result rather than deferring to the machine
The 4Ds are meant to be learned and practiced together; a team strong on Description but weak on Discernment ships confident garbage.
The three interaction modes
The framework separates what the human does from how much control they retain:
- Automation — the AI executes a task end-to-end with minimal human intervention. Fast, cheap, suitable for low-stakes or well-defined work
- Augmentation — human and AI work together, with the human in the driving seat. Appropriate for work where judgment, taste, or accountability matters
- Agency — the AI operates autonomously, pursuing goals, making decisions, and taking actions over extended horizons. Requires the strongest Discernment and Diligence, and usually additional technical guardrails
The right mode depends on stakes, reversibility, and the person's fluency. A task that is fine under Augmentation may be dangerous under Agency for the same operator. See ai-agent-infrastructure for the technical substrate that Agency-mode work depends on.
Why the vocabulary matters
Without shared language, teams collapse distinct questions ("should I use AI for this?" vs "how do I describe this to the AI?" vs "do I trust this output?") into a single fuzzy debate about "using AI." The 4Ds let teams name which dimension is weak. The interaction modes let teams name which mode a task should be in and notice when the actual mode drifts higher-autonomy than the task warrants.
This maps onto good-taste-as-competitive-advantage: taste is what powers Discernment, and Discernment is the scarce resource as AI output volume grows.
Prompt structure in practice
- Anthropic’s applied AI team frames effective Claude prompting as having six explicit elements, which reinforces that good prompting is a structured skill rather than a lucky phrasing problem
- The point is not prompt magic; it is making intent, constraints, and review criteria explicit enough that the model can act and a human can still inspect the result
- This is the practical face of Description and Diligence: the better the structure of the request, the better the downstream ability to judge the output
Expert prompts are explicit control contracts
Some power users now start with a durable expert instruction rather than a loose question. Marc Andreessen's current custom prompt tells the model to act as a world-class expert, answer in complete detail, and stay highly specific. The exact wording is less important than the pattern: role, depth, and output expectations are being encoded up front, which is Description turned into a control contract.