GenAI Enablement: AI for Everyday Work

Syllabus outline

1. 01

   Foundations

   Large language models as a function from text to text: parameters, weights, and what training actually searches for. Tokenisation and one-token-at-a-time generation. Sampling and temperature, and why one prompt yields two answers. Hallucination as a consequence of the objective rather than a defect. Knowledge cutoff, the context window as working memory, the model as distinct from the app around it, and the two meanings of "agent".

2. 02

   Prompting

   The prompt as a brief rather than an incantation. Its five elements: the task, supplied context, the reader, output format, and a worked example. Zero-shot and few-shot. Iterative repair against restarting. The limits of prompting and what no wording can fix. Paired weak and strong prompts for each element, with both outputs.

3. 03

   Your Workbench

   Three interfaces to the same tool: graphical, command line, and API. Terminal essentials: working directory, navigation, and the few commands in daily use. Markdown as a portable working format. APIs distinguished from applications. Claude Code installed and run against a working folder, with the common first-run failures.

4. 04

   The Claude Ecosystem

   One subscription across several surfaces. Chat, Projects as persistent working context, and Artifacts as editable documents rather than downloads. Cowork for delegated jobs, Claude Code for local folders, Claude Design for visual output. Choosing a surface by task. Flat-rate plans against metered API pricing, and the equivalent products at OpenAI and elsewhere.

5. 05

   The No-Cost Path

   Local inference against hosted inference. LM Studio for running a model on a laptop without a terminal, and the hardware each model size demands. Kilo Gateway as a single key across providers, and configuring a command-line tool against it. Goose as a no-subscription alternative to Cowork. Cost per route compared.

6. 06

   Skills, Agents and MCP

   MCP connectors as access to other applications, and the permissions that access carries. Skills as encoded working method, contrasted with project instructions; ready-made skills, and writing one for a house format. Agents as a loop toward an outcome rather than a sequence of steps, and three characteristic failure modes.