Most data science applicants in 2026 are studying the wrong things and applying everywhere at once. Notes from reading hundreds of resumes: what to learn first, how to aim applications, and why referrals do the work.
Prompt engineering has stopped being the interesting part of LLM work. This is about context engineering, caching structured context in transients, and checking that the model’s reasoning holds up.
Zero-shot prompts make the model guess your coding standards. Feeding it two or three working examples from your own code gets output that matches your prefixes, namespaces and security patterns.
Anthropic’s January 2026 report puts the correlation between prompt sophistication and response sophistication at 0.925. Prompt engineering did not die, it moved from magic phrases to knowing your domain well enough to frame the problem.
WordPress.com switched its MCP server to OAuth 2.1 with PKCE, so AI agents no longer need long-lived tokens pasted into a config file. What the code verifier protects against, and what an authorized agent can do on your site.
Low-tech logistics companies do not need AGI, they need to stop counting stock with a pen. Voice-driven inventory counts in n8n, damage reports generated from a photo, and a WordPress REST endpoint that takes the data in without breaking WooCommerce stock.
Perplexity and BLEU tell you nothing about whether an automated customer journey actually goes anywhere. The CDP framework scores continuity, deepening and progression against a taxonomy of journey stages instead.
LLM-as-judge asks a model that hallucinates to catch hallucinations. Displacement Consistency scores the vector between a question and its answer against grounded neighbors instead, with no source document and no second model call at inference time.
After 14 years I moved most of my work out of the IDE and into the terminal: Claude Code in Warp, a Git worktree per agent so they stop colliding, and stored slash commands so I am not retyping the same prompt.
AI can write a WooCommerce feature in seconds, and it usually writes it as one 150-line God function with a hard-coded API key. This is what that debt costs and how to refactor it into something maintainable.