Jargon

Jargon, explained

AI comes with a lot of shorthand. Here is every term used on this site, in plain words. If you have used ChatGPT or Claude, you already know enough to read all of it.

AI basics

The words that come up whenever AI is involved, whatever you are building.

AI model (LLM): LLM stands for large language model. It is trained on a vast amount of writing and code, and produces an answer by predicting what comes next, one small piece at a time. ChatGPT, Claude and Gemini are apps built on top of models like these. The same kind of model can write code, which is what Xanadu uses it for.; See also: Tokens, Frontier model, Open model
Tokens: A token is a chunk of text, roughly three-quarters of an English word. Hosted AI services count every token you send and every token they write back, and bill or limit you by that count. "I ran out of tokens" means hitting that limit. Running a model on your own computer has no per-token cost.; In everyday terms: Minutes on a phone plan: the meter runs on every word in both directions.; See also: Local model
Frontier model: The newest, largest models from the leading AI labs. They handle the hardest problems best. You reach them over the internet, your data goes to the provider, and you pay for every token. Xanadu uses one only when a task has beaten everything cheaper.; See also: Tokens, Open model, The ladder (tiers and escalation)
Open model: Also called an open-weight model. Companies such as Google, Alibaba and Mistral release models that anyone can download and run. They are smaller than frontier models, but the best of them are now good enough for most everyday coding work.; See also: Local model, Frontier model, Parameters (27B)
Local model: An open model downloaded and run on hardware you control: a laptop, a desktop, or a server you rent. Your code and data never leave it, it works without an internet connection, and the cost is the hardware and electricity rather than a meter per token. Xanadu does nearly all of its work this way, on a MacBook.; In everyday terms: The crew on site, rather than a subcontractor across town you pay by the hour.; See also: Open model, Tokens, Hardware you control
Hardware you control: Your own laptops and servers, or machines you rent by the hour that run only your software. For a company this is often the deciding point: proprietary code and customer data never pass through a third party’s AI service.; See also: Local model, GPU
Parameters (27B): Parameters are the numbers a model learned during training. Their count is the usual shorthand for size: a 27B model has 27 billion of them and fits in the memory of a well-equipped laptop. Frontier models are far larger and need data centres.; See also: Open model, GPU
MLX: A machine-learning framework from Apple, built for the chips in current Macs. It is one of the ways a Mac can run a sizeable open model at a useful speed.; See also: Local model
GPU: Graphics processing unit: originally for drawing video-game graphics, now the workhorse of AI because it does many small calculations at once. A large model runs far faster on a GPU with plenty of memory.; See also: RunPod, Parameters (27B)
RunPod: One of several companies that rent GPU servers by the hour. Because it is just a rented machine running your own software, the code still stays on hardware you control for the length of the rental.; See also: GPU, Parallel workers, Hardware you control
Agentic assistant: A chat assistant answers questions. An agentic assistant takes actions: it opens files, edits code, runs tests and tries again, with you watching and steering. It is the step between autocomplete and a fully unattended loop.; See also: AI autocomplete (Copilot, Cursor), Autonomous development loop
AI autocomplete (Copilot, Cursor): Tools such as GitHub Copilot and Cursor suggest code as a developer works. They speed up typing and lookup, but the developer still decides every change. This is what most people mean by "using AI to code", and it is not what Xanadu does.; See also: Agentic assistant, Autonomous development loop

How Xanadu works

The parts of the system that builds the projects on this site.

Autonomous development loop: Instead of a person prompting an AI one request at a time, the loop takes a whole plan and runs unattended, often overnight. Each task is written, tested and either accepted or retried with the reason it failed. People come in at the start (what to build) and when something genuinely needs a decision.; In everyday terms: A general contractor running a job from the blueprints, calling you only when a decision is yours.; See also: Harness, Roadmap, Gate
Harness: An AI model on its own only answers what it is asked. A harness wraps it with everything else a job needs: which task is next, what files and instructions to show the model, how to check its work, what to do on failure, and what to record. Most of what makes Xanadu reliable lives in the harness, not the model.; In everyday terms: The general contractor: not the one swinging the hammer, but the one who decides who does what and checks it.; See also: Autonomous development loop, The ladder (tiers and escalation), Gate
Roadmap: A plain text file listing everything to build, in order, grouped into phases. Each line is one task precise enough to check. Xanadu ticks tasks off as they pass, so the roadmap is always an honest record of what is done.; In everyday terms: The job schedule on a building site.; See also: Task, Backlog, Kit
Task: Xanadu’s tasks are deliberately small, usually one file and one change, because small changes are what an AI model gets right most often and what is easiest to check. A large feature becomes a few dozen tasks.; In everyday terms: A single work order: "hang the kitchen door", not "finish the kitchen".; See also: Roadmap, Gate
Kit: Before Xanadu builds a feature, a person (with a frontier model’s help) prepares a kit: a short specification, the core pieces written by hand and locked, the tests that define "done", and the roadmap of tasks. A good kit is most of the difference between a run that finishes and one that wanders.; In everyday terms: Blueprints, the materials list and the inspection checklist, ready before the crew arrives.; See also: Roadmap, Locked contract, Tests
Locked contract: The shapes of the data and the key connections between parts are written by a person and marked locked. The models build around them. This stops a model from quietly inventing its own version of something that other tasks depend on.; In everyday terms: Load-bearing walls: everyone builds around them; nobody moves them.; See also: Kit
Planner, executor, validator: The planner turns the roadmap into the next task and what the model needs to see. The executor is the AI model writing the change. The validator runs the checks (static analysis and the task’s tests) and decides whether the change is accepted.; See also: Autonomous development loop, Static analysis, Tests
Tests: Automated checks that run the code and compare what it does with what it should do. In Xanadu the tests are written before the code, so they define what "done" means, and an AI model can’t talk its way past them.; See also: Gate, Test safety net
Gate: Every task names its gate: usually a specific test, sometimes a code check. The harness runs it on every attempt. Only a change that passes is merged into the project; anything else is retried with the failure explained.; In everyday terms: The inspection: the work isn’t signed off until the inspector passes it.; See also: Tests, Merge, First-pass success rate
Static analysis: Tools that read code and flag errors, missing pieces and risky patterns before it ever runs. It is fast and catches a large share of an AI model’s slips, so it runs on every attempt alongside the tests.; See also: Tests, Gate
Merge: Each task is worked on separately; when it passes its gate it is merged into the main copy of the project, with a record of what changed and why. A failed attempt is simply discarded.; In everyday terms: Signing off a finished job.; See also: Gate
The ladder (tiers and escalation): The models are arranged in tiers from small and fast to large and capable. A task starts at the bottom. If it fails repeatedly, the harness escalates it one rung up. The paid frontier model sits at the top and is reached only when every local tier has failed, so expensive compute goes only to the problems that need it.; In everyday terms: The apprentice tries first; the master tradesman is called only when needed.; See also: Local model, Frontier model, Harness
Parallel workers: When a roadmap is large, Xanadu runs several workers side by side, each on its own task and its own copy of the project, merging as each one passes. More workers means the same plan finishes in less calendar time.; See also: RunPod, GPU
BAD_PATTERNS: When the same kind of error shows up more than once, it is written down as a pattern. From then on, any change matching it is rejected straight away with the known fix, so the loop doesn’t spend attempts learning the same lesson again.; In everyday terms: The site’s "we’ve been burned by this before" checklist.; See also: Rule promotion
Rule promotion: Repeated corrections are collected and, once they prove themselves, promoted into the shared rules the models are given on every task. The system gets better at avoiding a mistake instead of just fixing it each time.; See also: BAD_PATTERNS, Velocity log
Velocity log: Every attempt is logged. Over time the log shows which kinds of task are quick and which are hard, and that shapes how future work is split up and estimated.; See also: First-pass success rate
First-pass success rate: A measure of how well tasks are written and how capable the models are. A high rate means less time and compute spent on retries.; See also: Gate, Velocity log
MCP: Model Context Protocol: an open standard, introduced by Anthropic, for connecting AI assistants to other software. Xanadu offers one, so an assistant such as Claude can launch a run, check progress or stop it on request.; In everyday terms: A universal power socket: any assistant that has the plug can use any tool that has the socket.; See also: Agentic assistant
Spec (specification): A specification describes the rules precisely enough that two people would build the same thing from it. "Normative" means it is the authority: if the code and the spec disagree, the code is wrong. Writing the spec and its tests first gives Xanadu a target it can check itself against.; See also: Kit, Tests

Building and shipping software

Product and engineering terms used in the consulting offers.

Backlog: Everything a team intends to build but hasn’t, ordered so the most valuable work comes first. In Xanadu the backlog is the unticked part of the roadmap.; See also: Roadmap, Agile
Agile: Work is broken into small pieces and delivered in short cycles (sprints), so what gets built next is decided by what was learned from the last piece. Xanadu’s small tasks and frequent merges are the same idea, applied to the code itself.; See also: Sprint, Backlog
Sprint: A team commits to a set of work for the sprint, delivers it, reviews what happened, and plans the next one.; See also: Agile
SWOT: A simple grid: what is good and bad inside the product (strengths, weaknesses), and what could help or hurt it from outside (opportunities, threats). Useful at the start of a project for deciding what to build first.
Cycle time: One of the clearest measures of a team’s speed: the time from "work started" to "in users’ hands". A pilot measures it before and after, so the effect is a number rather than an impression.; See also: Throughput, Pilot
Throughput: Measured per week or per sprint. Together with cycle time it shows whether a team is really shipping more.; See also: Cycle time
Escaped defects: The measure that keeps speed honest: shipping more is no help if more bugs reach customers. A pilot tracks it alongside cycle time and throughput.; See also: Cycle time, Tests
Pilot: Rather than a promise, a pilot runs the loop on one team’s actual backlog for about six weeks, with the same measures taken before and after. The team keeps the setup and the numbers whatever they show.; See also: Cycle time, Throughput, Escaped defects
Test safety net: Before modernising old code, tests are written that pin down exactly what it does now, quirks included. Then every change, by a person or by the loop, has to keep those tests passing, so nothing the business relies on breaks unnoticed.; See also: Tests, Gate
Commit: Code is kept in a version-control system (almost always Git) that records every change. Each recorded change is a commit, with who made it, when, and why. Counting commits is a rough measure of activity; reading them tells the story of a project.; See also: Merge
PM (product manager): A product manager talks to customers, sets priorities and writes the roadmap; engineers build it. A "technical PM" also understands how it is built well enough to shape it.; See also: Roadmap, Backlog
IP address: Every device that connects to a website sends one, so that the reply can find its way back. Web hosts record them in standard logs; they can hint at a rough location but do not name a person.
Git: Git keeps the full history of a project as a series of commits, and lets work happen on separate copies (branches) that are merged back when ready. Xanadu keeps its progress in Git, which is why a run can be stopped and picked up again.; See also: Commit, Merge

The tech behind the projects

The languages and services named on the project pages, in a sentence each.

Flutter: Instead of writing an iPhone app and an Android app separately, a Flutter app is written once, in Dart, and runs on both, and in a browser.; See also: Dart
Dart: A modern language designed for building apps. Most of the phone apps on this site are written in it.; See also: Flutter
Firebase: Apps need somewhere to keep accounts and data. Firebase provides those pieces as a service, so a small team can ship an app without building and maintaining servers.; See also: Firestore, Cloud Functions
Firestore: A cloud database from Google. Changes on one phone appear on the others, and an app keeps working without signal and catches up when it reconnects.; See also: Firebase
Cloud Functions: For the jobs an app can’t do on the phone itself (calling an AI model with a secret key, sending a notification), a function runs in the cloud on demand. There is no server to keep running in between.; See also: Firebase
Next.js: Next.js takes care of pages, links, loading speed and search-engine friendliness, so the work goes into the product rather than the plumbing.; See also: React
React: React lets a page be built from components (a button, a card, a chart) that update themselves when the data behind them changes. It is the most widely used way to build web interfaces.; See also: Next.js, React Native
React Native: The same idea as Flutter, from the React world: write the app once and ship it to both app stores.; See also: React, Flutter
TypeScript: TypeScript adds types (this is a number, that is a date) to JavaScript, so tools can spot mismatches as the code is written. That makes it a good fit for AI-written code, because mistakes surface immediately.; See also: JavaScript, Static analysis
JavaScript: Anything interactive on a web page runs on JavaScript. With Node.js it also runs on servers.; See also: TypeScript, Node.js
Node.js: Node.js lets the language of the web power back ends, scripts and command-line tools too.; See also: JavaScript, TypeScript
Python: Python is popular for automation, data and AI because it is quick to write and has a library for almost everything.
Shell scripts: The glue of software work: start this, wait for that, restart it if it stops. Xanadu’s supervisor, which keeps a run going overnight, is a shell script.
Tailwind CSS: Instead of a separate style sheet, each element carries names like "rounded" or "text-gray-600". It keeps a site’s look consistent and is easy for AI models to get right.
Vitest: Vitest runs a project’s tests in seconds, which matters when a loop runs them on every attempt.; See also: Tests
Drizzle: Drizzle describes the database’s tables in TypeScript, so a mistake in a query is caught before it runs.; See also: TypeScript
Anthropic API: Chat apps are for people; an API is for programs. Xanadu uses Anthropic’s API to reach Claude when a task has beaten every local model.; See also: Frontier model, Tokens, The ladder (tiers and escalation)