Prompt Engineering –> Context Engineering –> Intent Engineering (twitter.com)

The piece argues the industry is moving beyond simple prompt engineering toward two broader, more robust disciplines: context engineering and intent engineering. Prompt engineering—crafting single-shot or few-shot inputs to coax desired outputs—worked for early LLMs but is brittle at scale. Context engineering expands this work to manage all the surrounding signals an LLM consumes: system prompts, retrieval-augmented knowledge, memory, tool calls, chain-of-thought traces, and context-window management. Intent engineering goes further by explicitly modeling and operationalizing user goals, mapping ambiguous requests to structured intents, policies, and action plans that agents or pipelines can execute reliably. This shift matters because it addresses the major failure modes of current systems—hallucination, inconsistency, and fragility in multi-step tasks. Technically, it emphasizes embeddings and RAG for grounding, long-context architectures and chunking strategies, intent classification and slot-filling models, formal intent representations (schemas/ontologies), and closed-loop verification (execution traces, tool outputs, and feedback loops). For ML practitioners and product teams, the change demands new tooling: context management layers, intent catalogs, observability for prompts+context, automated testing, and alignment/safety checks. The result should be more reliable, explainable, and automatable LLM-driven applications rather than brittle prompt hacks.