At time of writing, AI needs humans to do anything useful and there is a big difference between the best and worst employees at using AI.
We all need to build the skill of using AI. Skills are not tools—you want to build up a generalized set of skills not become a consumer for a bunch of tools that become obsolete as foundation AI models get better.
What skills you need to AI in 2026:
- Research - get up-to-speed quickly about any topic at any depth using AI to survey and ingest an large amounts of material then compress it down to answers (and more questions!). The human-in-the-loop is necessary for raising new lines of inquiry, asking new questions, and fact checking results.
- Coding - generate reasonable code at fast pace while maintaining high quality. The key skill to work on right now is splitting up work so it can be parallelized, pre-work to guide coding agents to solve the right problems, and establishing automated controls that prevent bugs and quality issues. Code is still primarily written for humans.
- Evals - write representative and objective evaluation criteria that makes it clear if generative AI output is getting better or worse. The bar is currently 500 high quality examples that needs to be curated carefully (you can always generate more synthetic data from there).
- Fine tuning (LoRa) - maybe.
- Agents use - I don’t know that this is a skill yet, but there is some growing area of technology decision making around deciding what agents with which scope make sense where and overall to accomplish a collective goal. Should it be an AI SDR Agent or should it be a Lead Enrichment Agent + Email Copywriter Agent + CRM Data Agent with some orchestrator on top? AI as employee is an attractive analogy, but this is far too wide to be useful at the moment.
Currently not worth it:
- Training new foundation models (you will make more progress by standing still and letting foundation labs do their thing).
- Prompt engineering (you don’t need little wording hacks anymore—built-in tool calls, longer context windows, and better instruction following obviate the need for it. However, context engineering and making sure the right data ends up in the prompt is much more important for AI applications and agents).
- RAG (this is built in to enough tools at this point that having the skills to build out your own RAG pipeline isn’t so critical anymore).