The AI Fluency Framework

The phrase “AI literacy” gets used a lot. It means almost nothing, because it conflates knowing what an AI is with being any good at using one. Fluency is the more useful word. Fluency means you can do work with an AI that you couldn’t do without one, reliably, in your own context.

The framework has three layers.

The four competencies

Fluency is the intersection of four skills, none of which can substitute for the others.

Delegation. Knowing what to give the AI in the first place. The fluent user has internalised which kinds of tasks return signal and which return slop, and matches the tool to the task.

Description. Specifying the work clearly enough that the model can actually do it. This is the prompt-engineering skill, but the deeper version of it, spec-writing, not magic-incantation hunting.

Discernment. Reading the output critically. Catching hallucinations, recognising plausible-but-wrong answers, knowing when to push back and when to accept.

Diligence. Closing the loop. Verification, attribution, the small disciplines that turn a draft into a defensible artifact.

A team that’s strong on description but weak on discernment ships confident garbage. A team that’s strong on discernment but weak on delegation never starts.

The three interaction modes

Fluent users move between three modes deliberately:

Automation, the AI does the task end-to-end, the human checks the output. Best for high-volume, low-stakes work with clear quality criteria.

Augmentation, the human and the AI co-produce the work in turns. Best for thinking work where the human’s judgment is the bottleneck.

Agentic, the AI takes multi-step actions in the world (tools, files, APIs), the human supervises the trajectory. Best for processes where the steps are well-defined but tedious.

Most people get stuck in one mode and try to use it for everything. Fluency is moving between them on purpose.

The prompt formula and six techniques

The base formula: role + task + context + format + constraints. Six techniques layer on top, chain-of-thought, few-shot, self-critique, decomposition, structured output, and tool use. Each is a tool, not a religion. Fluent users pick the technique that fits the failure mode in front of them.

The whole framework exists because most “AI training” teaches the techniques without teaching the underlying competencies. People walk out of those sessions able to copy a prompt template and unable to recover when the template doesn’t work. The competencies are what let them recover.