Day 1: How a Model Learns (Teachable Machine Live), Context Engineering, CO-STAR, Artifacts, Connectors, and Claude Cowork Running a Daily Stand-up

PROCEDURE: Google Teachable Machine -> Image Project -> Standard image model -> classes 'cats' and 'dogs' -> upload labelled images from a Kaggle dataset (100 per class) -> default epochs and batch size -> Train -> test on an unseen dog: 98% cat. Retrain with 500 per class: correct on both. The point, made without a line of code: training-data volume and quality decide accuracy. Framed against narrow AI (spam filters, credit scores, Face ID) versus generative AI's next-word prediction, and Google-as-retrieval versus AI-as-'context engine' - factual only as far as its references.

The whiteboard rule: sentences are cut into tokens 'like a cake', each token gets a numeric address with related concepts nearby, attention weights the words that matter, and the model predicts. Context windows grew from GPT-3's ~2-4K tokens (six pages, 2022) to file attachments (~64 pages, late 2024) to today's book series - but models remember the start and the latest turns best and can hallucinate on facts buried in the middle. 'Capacity is not equal to capability.'

Five kinds of context: identity (role or persona), world (situation and audience - 'presenting to a CEO'), task, example (reference material), constraint ('three slides'). A prompt is a command: be specific (the coffee-ordering parody), state expectations, don't mix instructions. Three habits: edit the prompt instead of piling on follow-ups; a new chat per topic; add output constraints such as 'no commentary, just output'. System prompt versus user prompt via Iron Man - the character brief versus the scene - demoed by generating a Tony Stark persona from a screenshot and watching Anthropic's guardrails object until it became 'a character from a superhero series'.

CO-STAR = Context, Objective, Style, Tone, Audience, Response - 'around a couple of years, that's a big deal'. PROCEDURE: create a project (Claude) or custom GPT with no instructions; paste the framework screenshot and ask 'convert this into a system prompt for my prompt-engineer assistant'; put the result in the project's instructions; from then on drop any task in and get a full CO-STAR prompt to run elsewhere. Model choice: Haiku fast, Sonnet balanced and good at code, Opus most capable. The lab later builds the same project with the framework file as knowledge.

Artifacts (Claude) and Canvas (ChatGPT) build interactive HTML/React front ends from a prompt. PROCEDURE from the lab: claude.ai -> Settings -> Capabilities -> enable the three visual toggles -> new chat -> a detailed prompt (an expense tracker) -> choose the model -> test (add entries, categories, weekly and monthly views) -> Publish artifact to the team or the web. Limits stated: front end only, no database, not sellable in app stores - micro-tools for yourself and your team. A Chat has no memory across chats; a Project keeps instructions and knowledge across every chat inside it (one notebook per subject) and spends tokens more efficiently.

Connectors let the model act on Microsoft 365, Canva, Gmail, Calendar, Zoom, Slack, CRMs - the brain in a jar gets hands. Named plainly: 'the technology it uses underneath is called MCP.' PROCEDURE: enable the Excalidraw connector, ask for the session as a flowchart, get a link with the canvas filled. Advice: connect first the tools behind your most hated recurring tasks (calendar, email).

PROCEDURE: new Cowork task -> pick a local folder (his 'ADHD Shortcuts' app) -> Always Allow -> 'read the folder, summarise, propose v2 features' -> 'turn that into a two-week roadmap' -> create a scheduled task: daily 9 AM stand-up referencing the roadmap. It fails on first run (missing context); embed the roadmap text in the task prompt and it works. Then Claude for Chrome: 'open YouTube Music and play Daft Punk, then build an interactive artifact that reveals welcome to the agentic era', permission 'act without asking' - it errs, screenshots, self-corrects. Cowork and the extension are Pro features ($20/month pitched).

Separate learning AI (anonymised data, no worry) from using AI (enterprise plans with liability cover, local models, paid private cloud); becoming the internal AI champion who pushes for the enterprise plan is also career protection. Raghuveer's own two-year-refined account-level system prompt: no emojis, hype or filler; no mood mirroring; stop immediately after the information; 'restore independent high-fidelity thinking'; guard against model obsolescence - stay able to think if the tool vanished. A living document to build in Settings. OpenClaw was explicitly not taught: 'it will take access to literally everything on your system'.