why are there so many buzzwords in AI? (what to learn, full breakdown)

Prompt engineering. Context engineering. Harness engineering. Loop engineering. And now graph engineering. Five terms in about three years, and most people using them can't tell you what actually changed between them. In this video I break down all five with simple examples, no jargon, so a non-technical person can follow the whole thing. More importantly, I show you what each one was actually solving, because none of these appeared for fun. Each one exists because the previous approach hit a wall. A few things I say in here that people won't like: Persona based prompting ("you are an expert coder") has been studied and it does not improve output. We all did it anyway. Context engineering is still the base layer. Harness engineering, loop engineering and graph engineering are all extensions of it, not replacements for it. Loop engineering is token hungry by design. You are handing the agent permission to keep prompting itself until a goal is met, so you own the bill. Graph engineering is real and I have built with it, but I am not fully convinced by how it is being sold right now. You lose control over which sub agents get spawned, and it burns a lot of tokens to get there. CHAPTERS 0:00 Prompt to graph engineering in three years 0:44 The prompt engineering era (2022 to 2023) 1:17 Persona based prompting, and why it doesn't work 1:37 Few shot and one shot prompting 1:56 The 4,000 token problem 2:17 Context windows explode, Cursor and MCP arrive 3:02 Context rot: the hidden ceiling on big context windows 3:34 The context engineering era (still the most relevant one) 4:04 Opus 4.6 and the jump to long horizon tasks 4:33 What a "harness" actually is 4:58 Example: adding a returns and refunds workflow 5:50 Why compaction exists 6:23 Skills and the rise of software factories 7:07 Prompt vs context vs harness, in one image 7:38 Loop engineering 8:29 Example: an agent that fixes your CI/CD pipeline at 8am 9:43 The real difference: external verification 10:47 Loops burn tokens, budget for it 11:22 Graph engineering and dynamic workflows 12:08 One agent spawns many: the deep research example 12:44 Orchestrator, parallel paths, and separation of responsibility 13:38 The cons nobody mentions 14:07 The verifier agent 14:19 My honest take on graph engineering 14:35 Wrap up LINKS AND SOURCES Context rot study (Chroma Research): https://research.trychroma.com/contex... Context rot toolkit on GitHub: https://github.com/chroma-core/contex... Effective context engineering for AI agents (Anthropic): https://www.anthropic.com/engineering... Long horizon task graph (METR, Measuring AI Ability to Complete Long Software Tasks): https://metr.org/blog/2025-03-19-meas... METR paper on arXiv: https://arxiv.org/abs/2503.14499 Persona prompting study (When "A Helpful Assistant" Is Not Really Helpful): https://arxiv.org/abs/2311.10054 Claude Opus 4.6 announcement: https://www.anthropic.com/news/claude... Dynamic workflows in Claude Code (docs): https://code.claude.com/docs/en/workf... Dynamic workflows announcement: https://claude.com/blog/introducing-d... A harness for every task: https://claude.com/blog/a-harness-for... Agent Skills (docs): https://platform.claude.com/docs/en/a... Anthropic open source skills repo: https://github.com/anthropics/skills Model Context Protocol: https://modelcontextprotocol.io Introducing MCP (Anthropic): https://www.anthropic.com/news/model-... Cursor: https://cursor.com FIND ME HERE LinkedIn: LinkedIn: parikshitpruthi Instagram: Instagram: parikshitpruthi YouTube: @parikshitpruthi If you build with agents, tell me in the comments which of these five you are actually using in production. I read every comment. Subscribe if you want more breakdowns like this one. #GraphEngineering #ContextEngineering #PromptEngineering #AIAgents #ClaudeCode #AgenticAI #LLM #SoftwareEngineering