PAGES40
WORDS53,529
SOURCES298
READ TIME86 min
TYPEHUB
CHECKED25 AUG 26
Everything on this site about assembling things with AI, ordered from the decision to build through to handing it over.
◈ THE THING THIS HUB ADDS
Most guides ask how do I build this. The more useful question is who a given setup suits — because the same technique is right for one kind of work and wrong for another.
A prototyper needs loose loops and throwaway scaffolding. A maintainer needs permission tiers and an audit trail. Handing either one the other’s setup makes them slower. The five roles are on the archetypes page.
Start here
If you have not built anything yet.
GUIDE · 14 MIN · 22 SOURCESFree Alternatives: Run AI on Your Own ComputerHow to run AI free on your own computer with Ollama or LM Studio: the memory you need, the steps, which models to try, and what each licence allows.GUIDE · 12 MIN · 28 SOURCESPrompting Myths: Advice That Stopped Being True (or Never Was)Common prompting rules, from “think step by step” to “8K masterpiece”, checked against what OpenAI, Anthropic, Google and image vendors actually publish.GUIDE · 11 MIN · 23 SOURCESHow to Test Whether a Prompt Change Actually HelpedA simple way to test whether a prompt change helped: fixed test inputs, one change at a time, several runs each, and “better” defined before you look.GUIDE · 4 MIN · 4 SOURCESPrompt Caching: The Biggest Discount Nobody Turns OnCached input costs a tenth to a quarter of the normal rate at major vendors, but a cache write costs extra. Where it breaks even, and why prompt order matters.EXPLAINER · 3 MIN · 3 SOURCESMeta’s Developer Stack: Muse Spark, Code, Image and the Model APIMeta’s developer stack: Muse Spark on the Model API, Muse Code, an image model and transcription. What each does, how each is priced, and what is still unclear.EXPLAINER · 4 MIN · 4 SOURCESWhen AI Models Retire, and How Much Warning You GetOpenAI promises at least six months. Anthropic promises sixty days. Both published policies, the real gaps between announcement and shutdown, and what to do.COMPARISON · 4 MIN · 3 SOURCESLLM API Prices: OpenAI, Anthropic, Google and xAITwelve models, four vendors, one table — every figure read off the company’s own pricing page. What they agree on, and the two changes already scheduled.COMPARISON · 4 MIN · 6 SOURCESWhat the OpenAI API Actually CostsFour things decide your OpenAI bill and only one is the headline price: the model ladder, the long-context cliff, caching, and the spread between service tiers.GUIDE · 3 MIN · 2 SOURCESHow to Build for Meta Ray-Ban Display GlassesMeta opened the display to developers in May 2026: a native SDK, or plain HTML deployed by URL. The two paths, the wristband input, and the preview limits.EXPLAINER · 4 MIN · 2 SOURCESWhat Is Gemini Spark? Google’s Background Agent, ExplainedGemini Spark runs on a schedule, not on a prompt. What Tasks, Skills and Schedules do, which apps it reaches, who can use it, and what Google will not say.COMPARISON · 4 MIN · 4 SOURCESGemini vs Grok API Prices, Side by SideThe published per-token prices, read off both vendors’ own pages. Including the Gemini rate that doubles on 1 January 2027, and why Meta is not in the table.REFERENCE · 4 MIN · 2 SOURCESEvery Google AI Tool in 2026, and What It Is Called NowWhisk, ImageFX, NotebookLM, Gemini CLI, Vertex AI and Imagen 4 were all renamed, merged or shut down in 2026. The current list, with dates and sources.GUIDE · 4 MIN · 3 SOURCESHow to Run Muse Glimmer LocallyMeta’s 30B open agent model needs about 24 GB of GPU memory in its quantised form. The published sizes, the routes Meta names, and what is not confirmed yet.EXPLAINER · 3 MIN · 3 SOURCESGoogle AI Studio vs Gemini Agent PlatformOne is the free console for trying things; the other is what Vertex AI became on 23 April 2026. How to choose, and the privacy line most people miss.EXPLAINER · 2 MIN · 1 SOURCESGemini CLI Is Gone: Moving to Antigravity CLIGemini CLI stopped serving Google AI Pro, Ultra and free Code Assist users on 18 June 2026. What replaced it, what carried over, and who was unaffected.EXPLAINER · 4 MINPrompting 101: The Complete Beginner's Guide to AI PromptingNew to AI prompting? Learn what a prompt is, why specificity matters, and the exact structure that separates amateur AI outputs frEXPLAINER · 3 MINWhat a Repo Is, and How Repos Actually WorkA repository is a folder with a memory. Explained plainly, with sources named and dated.EXPLAINER · 4 MINWhat AI Skills Are, and How They WorkSkills are packaged know-how an AI loads on demand; tools are the functions it can call; MCP is the open socket…
Before you build
Two of these questions should sometimes stop you.
THE CRAFT · 3 MIN · 1 SOURCESThe Agent BriefA copyable structure for briefing an agent — objective, scope, done-condition, stop-conditions, evidence — with why each part is there and what to cut.THE CRAFT · 5 MINAutomationsChains that run without you. Reversibility as the design axis, idempotency, what to log, where the human checkpoint goes, and whenTHE CRAFT · 2 MINPrompt, Context, Skill, or Fine-TuneFour ways to shape model behaviour at very different costs. Most fine-tuning problems are context problems, and a fine-tune is lost when its base model retires.THE RECORD · 4 MINWHEN NOT TO USE AINine cases where reaching for AI makes the outcome worse, and how to tell in advance.
Who each setup suits
The same technique is right for one kind of work and wrong for another.
Agents
What they are, how they fail, and how to bound them.
THE CRAFT · 6 MIN · 2 SOURCESPrompting AgentsBriefing something that acts: the specification, the six ways agents fail, the dry run, and designing work so you can actually cheTHE CRAFT · 3 MINPermission Tiers for AgentsFive permission levels for agents, set per task by how hard an action is to undo, not by how far you trust the agent, and why grants should end with the task.THE CRAFT · 2 MIN · 1 SOURCESMulti-Agent: When It Helps and When It Does NotWhy chaining agents cuts reliability — compounding error, lossy handoffs, failures you cannot trace — and the cases where splitting the work earns its cost.THE CRAFT · 2 MINDiagnosing a Run That Went WrongThe agent finished and the output is wrong. How to find the first wrong step in the transcript rather than its consequences, and the failure signatures to spot.DATED · 9 MIN · 15 SOURCESAI AGENTSAgents act instead of answering. What changed, when, and what it actually means.
Skills and context
Reusable instruction, and what it costs to trust someone else’s.
THE CRAFT · 2 MINSkills, Repos and MCPWhat these three actually are, explained without assuming you write code.THE CRAFT · 3 MIN · 1 SOURCESSkills Are a Supply ChainInstalling a skill loads someone else’s instructions into your agent. An audit of 3,984 published skills found 13.4% had a critical flaw and 76 were malicious.THE CRAFT · 3 MIN · 2 SOURCESContext EngineeringWhat surrounds a prompt — the material, the history, the tools, the memory — now matters more than its wording. Four rules that hold as the tooling changes.THE CRAFT · 3 MINReusable Context Beats Reusable PromptsPrompts are tuned to one model and expire with it; context about your own work does not. How to turn a prompt library into one that survives model releases.
Site data: each card restates the page it links to, where the figure is sourced and quoted; cards checked against those pages 23 Sep 2026.
Loops and review
Which iteration converges and which quietly drifts.
THE CRAFT · 3 MIN · 2 SOURCESLoops That Improve and Loops That DegradeSome AI iteration converges; some only polishes its own output. The difference is whether new information enters each pass, and which common loops do.THE CRAFT · 2 MINThe Review Loop, In OrderReview AI output in order: frame, facts, structure, language. Each pass can undo the work below it, so starting with the wording means doing it twice.EXPLAINER · 5 MINThe Iterative Workflow: How to Refine AI Output Instead of Starting OverThe 4-round iterative AI workflow, img2img refinement, and when to cut your losses. Professional AI output rarely comes from firstTHE RECORD · 3 MIN · 1 SOURCESAuditing AI-Generated WorkChecking everything is not a strategy. Four risk tiers with a different check for each, the signals worth looking for, and a monthly review of your own use.THE RECORD · 3 MINThe Cost of Checking Too MuchEverything says check more. Almost nothing says where checking stops paying. Review theatre, uniform attention, and the failure no
Cost and handover
What it burns, and what you owe whoever inherits it.
THE CRAFT · 3 MIN · 1 SOURCESWhat You Are Actually BurningYou pay for everything the model reads, re-sent every turn. Where the cost really goes, why agents multiply it, and the model choice that fixes most of it.THE CRAFT · 3 MIN · 2 SOURCESWhat You Owe a Client Whose System You AutomatedWhat you owe a client after automating their work: a handover document, disclosure, an off switch, a visible check for silent failure, and agreed limits.THE CRAFT · 3 MIN · 2 SOURCESAgent GuardrailsFour lines that belong in every agent prompt. Scope, negatives, injection defence and stopping conditions — the permission model t