Pick the Right Tool
Which AI for which job, when you need a chat versus an automation versus a little robot, when a decision model beats a writing model, and how to find the thing someone already built.Which model tier for which job, when a fixed automation beats an AI, when a decision model (pick / score / yes-no) beats an LLM, when you actually need an agent, and how to find what someone already built.Model selection by task, LLM vs. System One decision model vs. deterministic automation vs. agent loop, search-first sourcing, and token economics.
Carry on withCarry on withnext:
- 019 min
Which AI for Which JobWhich AI Model for Which Job (Pick the Tier, Not the Leaderboard)Model selection: a task-based comparison, not a leaderboardDoneDonedone
There is no best AI. There is the right one for reading a lab report, the right one for a thousand product descriptions, the right one for a quick question, and one that only sorts and never writes.There is no best model, only the right tier for the task: frontier for one-off hard reading, mid-tier for daily drafting, small for high-volume sorting, local for data that cannot leave, and a decision model when the answer is a label, a score or a yes/no. A ten-example test, rerun each quarter, tells you which.Frontier vs. mid-tier vs. small vs. local LLMs, plus a System One decision model for typed answers, mapped to cannabis tasks, with a test protocol you can rerun each quarter.
- 028 min
Chat, Automation, or Agent? A Simple DecisionChat, Automation, or Agent? What an Agent Loop Is and When It Is Worth ItLLM call vs. deterministic automation vs. agent loop: a decision matrixDoneDonedone
Most problems need a plain automation with no AI in it. Some need one AI call. Some need a routine that asks a sorter one question at a fork. Very few need an agent that decides things on its own.Most jobs are a fixed routine with no model in it (a deterministic automation). Some need one AI answer (a single LLM call). Some are a routine that asks a decision model one typed question at a branch point. An agent, the model sending itself the next message, is for the few jobs you cannot list step by step and can afford to undo.When to use a single completion, when to wrap it in a scheduled workflow, when code keeps control and asks a decision model one typed question at a branch, and when an agent loop is justified by the four criteria. Autonomy is one removed pause, and it has a price.
- 037 min
Someone Already Built It: How to Find ItSomeone Already Built It: Searching Repos, Templates and Registries FirstSearch-first: GitHub, n8n templates, skill directories and the MCP registry before you buildDoneDonedone
You do not have to solve every problem. Most of the time someone has already shared the answer for free. Here are five places to look, five questions to ask before you trust it, and how to make the AI do the searching.Most problems already have a free answer online: a repo on GitHub, an n8n template, a skill, an MCP server, or one of our giveaways. Ten minutes of searching, a five-minute check (last commit, issues, license, access, does it run), then adapt instead of build.A search protocol across GitHub, n8n templates, skills.sh, the official MCP registry and this course's giveaways; a five-question repo triage with the gh commands; fork-and-adapt over greenfield; and a prompt for model-assisted search with mandatory URL verification.
- 048 min
What This Actually CostsWhat This Actually Costs: Tokens In, Tokens Out, and Why Plugs Bill Every MessageToken economics: input vs. output, per-turn context cost, caching, batch, and a monthly estimate per workflowDoneDonedone
You pay by the word-chunk, and what the AI writes costs five times what it reads. A pasted spreadsheet is a few cents. A plug you leave switched on costs on every message. Here is how to estimate one month for one job.Input tokens are cheap, output tokens cost about five times more, and every connector you leave on re-sends its tool list each turn. A subscription is a flat fee; an API key is pay-per-token. Estimate a month for one workflow with the sheet, then pull the levers: smaller model, shorter context, cache, connectors off.Per-token pricing (input/output, cache read 0.1x, batch 0.5x), per-turn context re-send including tool definitions, subscription vs. API, decision models at $0.042/Mtok, and a CSV estimator for one workflow's monthly bill.
- 057 min
AI You Can Run on Your Own ComputerAI on your own computer: open models, when they win, and when they do notLocal inference: open-weight small models, trade-offs, and a working Ollama setupDoneDonedone
Some AI models can run on a computer in your own building, so nothing you type leaves it. That is great for some jobs and a bad deal for others. Here is how to tell which, and what it takes to try one this afternoon.Open-weight models (Llama, Qwen, Gemma, Mistral) can run on your own hardware, so data never leaves. When a local model wins (sensitive data, no internet, high volume of simple jobs), when it loses (hard jobs, no one to maintain it), and how to try one.Open-weight small language models run locally: privacy, cost at volume and offline operation versus lower capability, hardware limits and ops burden; quantisation in one paragraph; Ollama end to end with its OpenAI-compatible endpoint; how to decide with your own eval set.