The CLI-vs-MCP debate has covered tools the model already knows. It hasn't covered the harder case: agents needing domain expertise the model lacks. A domain CLI paired with a workflow playbook puts practitioner judgment — the kind the model can't generate on its own — into the agent's loop. The architecture is also a governance question: whose expertise travels through your tool.
Standards calcify. Accessibility governance spent 25 years learning what happens when compliance diverges from outcomes — AI safety is setting up the same dynamic, with those lessons available in advance.
LLMs produce technically-valid outputs that fail to serve disabled users. Context is the intervention - examining prompt engineering patterns that shift burden from users to systems.
How I used AI to build an offline transcription app that protects user privacy—and what it reveals about the democratization of software development.
Personal ethical stance.
See all tags.
