LLM Basics
Lesson 4 of 7

Task Planning & Decomposition

Prompt chaining, back in Prompt Engineering, assumed a person already knew the steps, summarize an article, then translate the summary, and simply ran them one at a time through separate calls. Task planning removes that assumption. Give the model nothing but an end goal, and ask it to invent the steps itself.

Predict before you look

The prompt below gives the model nothing but an end goal, launch a small internal newsletter, with no steps specified at all. Before you look, do you expect the plan it produces to be roughly what you'd write yourself, or noticeably different?

Given only the goal, launch a small internal newsletter for a 20-person team, and no hint at what the steps should be, the model produced a genuinely ordered plan, define the purpose, assemble a small team, pick a format and platform, build a content calendar, design a template, assign roles, test with a small group, then launch. Nobody told it any of that structure, it reasoned from the goal to a plausible sequence entirely on its own.

This is the real difference from prompt chaining. Chaining still needs a person to already know summarize comes before translate, the model just executes each step in order. Here the ordering itself, what comes first, what depends on what, is something the model worked out, which is a genuinely different kind of decision than anything earlier in this course asked a model to make.

What's actually different between prompt chaining and task planning, mechanically?