Ad hoc generator ad hoc generators
A throwaway script, rewritten in full each session, that produces a document and then disappears. To be distinguished from the stable artifact, which persists across sessions and is touched only when its logic changes.
AI agent agent
A system built around a language model that can plan and carry out several steps on its own to reach a given goal. Unlike a plain conversation, an agent can reach for tools — web search, code execution, file access — and chain actions together without a human stepping in at every stage.
Alignment
The set of techniques aimed at making an AI model act in line with human intentions and values. Anthropic pursues alignment research so that Claude stays helpful, honest and harmless.
Anchor key anchor
A stable identifier attached to every glossary entry, derived from the term itself. It serves as the anchor on web pages, so a link can point straight at the definition. Never typed by hand, and never changed once published.
Anthropic
The American AI company founded in 2021 and the creator of Claude. Anthropic specialises in AI safety research. Its approach aims at models that are reliable, interpretable and aligned with human values.
Anti-pattern anti-patterns, antipattern
A recurring mistake that experience has brought to light, recorded together with its consequence and its remedy. The Kit keeps a catalogue of them so the same mistake does not travel from project to project.
API
An interface for calling a model straight from a program, bypassing the chat interface. This is the route used to embed Claude in an automated workflow or an existing application.
Artifact stable artifact
The word covers two things. In the claude.ai interface, an artifact is a file or block of content produced by Claude and shown in a dedicated side panel: code, a document, a vector image, an interface component. In the Kit, a stable artifact is an executable tool that lives permanently in the project files and is not regenerated each session: validators, renderers, linters, helpers.
Artificial intelligence AI
The set of techniques that let computer systems carry out tasks normally requiring human intelligence: understanding language, reasoning, producing text or images. Large language models such as Claude belong to the generative AI category.
AST abstract syntax tree
A structured representation of a document in which every block — paragraph, table, image — is identified by its type. The Kit converts Word documents into web pages by walking this tree rather than the raw text, so that nothing is lost on the way.
Chain-of-thought chain of thought
The practice of explicitly asking for step-by-step reasoning before the conclusion. On analysis, diagnosis and multi-criteria comparison, answer quality improves markedly.
Chat conversation
The text dialogue interface to Claude, reachable at claude.ai. Each chat is a separate conversation with its own history and its own context window.
Chat history
The list of past conversations stored in the claude.ai interface. Claude can search it using its tools for retrieving earlier conversations.
Claude
Refers to the AI model developed by Anthropic, independently of the interface. Used in everyday and technical contexts where the AI suffix is dropped by habit. See Claude AI for the full product name.
Claude AI
The full name of the product developed by Anthropic, covering both the language model and the interface that comes with it (claude.ai). Claude AI is the main assistant used in this Kit for documentation, Node.js scripts and configuration files. To be distinguished from Claude, the model on its own.
Claude project
A dedicated workspace in claude.ai gathering conversations on one subject together with shared reference files. Each project has its own persistent memory.
Consumer project consumer projects
A project that uses the Kit without being part of it. It references the stylesheets where they already live, never copies them, and takes on Kit changes at the moment it needs them.
Context
Everything available to Claude during a generation: the system prompt, the earlier exchanges, attached files and the project files.
Context window
The maximum amount of text a model can handle at once, measured in tokens. Everything that comes before in the conversation takes up that window: the system prompt, the instructions, attached files and earlier replies. Once the limit is reached, the model can no longer see the oldest exchanges.
Control layers Layer 1, Layer 2, Layer 3
The three lines of defence protecting the production chain: the checks built into the functions themselves, the validators run around each generation, and the cross-cutting audits carried out in dedicated sessions. The earlier a mistake is caught, the less it costs.
Cover sheet cover page
An A4 PNG page at 300 dpi that opens the paper binder. It acts as a visual table of contents — sections, numbers, colours and documents — mirroring the physical dividers exactly.
Data module data modules
A small JavaScript file holding exactly one list: glossary terms, brand names, variable names, home-automation entities. It is loaded at the start of a generator script only if the project declares that it needs it.
Delta deployment delta pass
A publication that regenerates only those pages whose source document has changed since the last deployment, instead of rebuilding everything.
Divider dividers, tab
A coloured, numbered physical divider in the paper binder. Each one stands for a section, and its colour alone is enough to find it again without memorising numbers.
DXA twip, twips
The internal unit of measure in Word documents: one twentieth of a point, roughly 57 units per millimetre. Every indent, column width and spacing value in the Kit is expressed in DXA.
Embeddings semantic vectors
A mathematical representation of text as a high-dimensional numeric vector. Embeddings capture meaning and semantic relationships: sentences that are close in sense end up close as vectors. They are used above all in RAG systems to locate relevant passages in a document base quickly.
Fail loud early fail loud
The principle that an error should cause a loud stop as early as possible rather than slip by unnoticed. An error caught late costs disproportionately more, and a silent error is never fixed at all.
Few-shot prompting few-shot
The practice of showing one to three input-and-output examples before the real request, rather than describing the expected format in words. The model picks up the structure, tone and level of detail straight away.
File pair file pairs, pair
Two files that exist only together and always carry the same timestamp: a stylesheet and its reference document, a terms module and its glossary, a cover sheet and its image. Changing one means regenerating the other in the same session — otherwise documentation and code contradict each other.
Fine-tuning Finetuning
A training technique that adapts a general-purpose AI model to a specific domain or style by feeding it targeted examples. Fine-tuning permanently alters the model's internal weights. Not to be confused with prompt engineering, which acts only on the instructions given during a session.
Fingerprint custom property, custom properties
Two invisible properties written into every document produced: the stylesheet version used and the generation timestamp. They make it possible to tell whether a document is current — one without a fingerprint is turned away at web publication.
Gemini
The family of generative AI models developed by Google DeepMind. Gemini spans everyday use through to professional applications, and ranks among the main competitors of Claude and GPT-4.
Generation
The act by which the model produces a text reply, predicting token after token the most likely continuation of the context it was given. Every call produces a fresh generation, independent of earlier ones unless the whole conversation is sent again.
GPT-4
The large language model developed by OpenAI and, by extension, the GPT family of models. A frequent point of reference in generative AI comparisons and a direct competitor to Claude.
Grok
The family of generative AI models developed by xAI. Grok draws recent data from the stream of the X platform, to which it is directly connected, and uses a more forthright conversational tone than its competitors.
Grok AI
The full product name of the xAI assistant, interface included. Used where the distinction from the model alone matters. See also Grok.
Guardrails guard rails
Rules and filters built into the model or the system prompt to bound Claude's behaviour and head off harmful, inaccurate or inappropriate replies.
Hallucination
A case where an AI model produces information that sounds plausible but is factually wrong, invented or distorted. Hallucinations arise because the model generates statistically likely text rather than factually verified text.
Inference
The phase in which a trained model turns an input into a reply. It stands opposite the training phase. Every message sent to Claude triggers one inference.
Instance
An active occurrence of a model handling one particular conversation. Each tab or Claude session opens its own instance. Instances are independent of one another and do not share their working memory.
Inventory file inventory
A text file (.txt) listing every project file with its exact name and timestamp. Updated in every session that produces new files.
Iteration
A way of working in which a result is sharpened step by step through successive rounds with Claude. For anything complex, this is the natural and recommended approach.
JSDoc
A standardised comment placed above a function describing what it expects and what it returns. In the Kit, the JSDoc in the code and the reference document must always say the same thing.
Kit bug report bug report
A short document by which a consumer project reports Kit behaviour that cannot be fixed locally. It describes, reproduces and proposes — analysis and correction then happen on the Kit side.
Knowledge cutoff cutoff date
The date up to which the model's training data was collected. Claude knows nothing of events after it, unless a web search tool is available to it.
L10N i18n, internationalization, localization
The standard abbreviation for localization (L + 10 letters + N), usually named alongside i18n (internationalization, I + 18 letters + N). In the Kit, the L10N pattern is the arrangement that lets output be switched between languages: a translation table keyed by string, a setLanguage(code) call at the top of the script, and a getStrings() accessor used wherever a string depends on the active language.
Landing page home page
The single page at the root of a subsite, showing the colour-coded sections and the documents they hold. The digital counterpart of the binder's cover sheet without being a copy of it: the binder lists everything, the landing page aggregates.
Linter linters
A script that inspects a generator's source before it runs, spotting where the house rules have been broken. It halts generation rather than letting a faulty document be produced.
LLM Large Language Model, large language model
An AI model trained on vast text corpora to understand and produce natural language. Claude, GPT-4 and Gemini are examples of LLMs.
Model
A precise version of an AI system, identified by a name and a version number. Each model has its own capabilities, its own context window size and its own behaviour.
Multimodal
A model's ability to handle several input types in a single request — text, images, PDF documents. A screenshot, a diagram or a scanned plan is analysed just as text would be.
NBSP non-breaking space
A space at which the line is not allowed to break. It sits between thousands groups and before a unit, so that an amount is not torn apart at the end of a line.
Node.js
The JavaScript runtime the Kit uses to generate .docx documents and HTML pages. Every generator script is written in Node.js and relies on the docx library to produce structured Word files. The cover sheet PNG, by contrast, comes out of a Python renderer.
OOXML Office Open XML
The internal format of Word files: a .docx is in fact a compressed archive of XML files. That is precisely what lets a program read and modify a document without ever opening Word.
Open items next steps
A closing section of a session summarising the subjects left untouched, the decisions still pending and the avenues worth pursuing that came up along the way.
Paper binder binder
The printed form of a documentation project. Colour-coded, numbered dividers organise it so that a newcomer can pick the project up without opening a single electronic file.
PCL render constants, Printer Command Language
A block of named constants that fully describes the layout of a generated document — dimensions, colours, positions, margins, texts. A renderer reads them and produces the output file: same constants, same result, every time. The name is borrowed by analogy with HP's Printer Command Language and the PostScript family of page description languages, where a page is described through instructions rather than drawn by hand. The Kit emits no actual PCL commands — it is the principle that is borrowed, not the format.
Persistent memory Memory, long-term memory
The mechanism that lets Claude retain information across separate sessions. It draws important facts from conversations on its own, stores them as notes and feeds them back into later sessions. It is switched off in incognito mode.
Pipeline processing chain
A sequence of transformation steps applied to content. In the Kit the word stands above all for the body-text rendering chain — five passes in order: brands, glossary, variable names, home-automation entities, then URLs — and for the conversion of Word documents into web pages.
Predecessor and successor predecessor, successor
An element that a formatting function requires immediately before or after it. Forgetting one raises no error at generation — the document is produced as usual, but comes out wrong when opened.
Project files
A project's files as they live in its working tree: stylesheets, data modules, Registry, tools and produced documents. They form the project's memory, and every change departs from one of them.
Prompt instruction
A message or instruction sent to Claude to trigger a generation. How useful the reply turns out depends directly on how clear and precise the prompt is.
Prompt engineering
The craft of wording instructions so that a LLM gives the best possible answer. Proven techniques include few-shot prompting, chain-of-thought, role prompting and negative constraints — saying what you do not want as much as what you do.
RAG Retrieval-Augmented Generation
An architecture pairing a search engine over a document base with an LLM. Before answering, the system retrieves the most relevant passages and places them in the model's context. This lowers hallucinations and makes recent or proprietary data usable.
Registry Registry.js
The JavaScript file holding a project's active versions, timestamps and configuration flags. The single source of truth: where another document disagrees, the Registry is right.
Renderer
A script whose only job is to read the PCL constants and produce the output file. A renderer holds no hard-coded layout value: same constants, same result.
Role prompting
The practice of assigning the model a precise role to set its level of expertise, its vocabulary and its register. A role that stays generic adds nothing: naming the domain and the audience precisely is what makes the difference.
Sequential delivery
The Kit's working rule whereby each file is generated, validated and presented to the user before the next one is started.
Setter setters
A function called at the top of a generator script to hand it a set of data: glossary terms, brand names, variable names, home-automation entities. If the setter is left out, the matching rendering is silently switched off — the script runs without error and the document comes out incomplete.
Spread spread operator
In JavaScript, the operator that unfolds an array returned by a function and inserts its elements one by one. Some Kit functions return an array: forgetting it produces a stray element that breaks conversion to a web page.
Stylesheet style sheet
In the Kit, a JavaScript module (.js) gathering all visual constants along with the functions that build paragraphs and tables. Three of them produce Word documents — General, Glossary, YAML — and a fourth produces the web pages.
Subsite subsites
A website of its own for one project, published under its own subdomain and generated from the same documents as the paper binder.
Surgical patch surgical patches
A targeted change to an existing document, confined to what was explicitly asked for, without regenerating everything. Reserved for cases where a full regeneration would be out of proportion.
System prompt
Instructions set before a conversation begins and invisible in the thread itself. They define how the model behaves: its role, its tone, its rules and its limits. They take up the context window just as messages do.
Temperature
The setting that governs how creative or how deterministic generation is. A low temperature yields predictable replies; a high one favours variety.
Template outside the Kit templates outside the Kit
A document form of a project's own, created outside the forms the Kit provides, when the subject demands it. The right is declared in the Registry and bears on forms, never on rules: the project keeps the Kit's reflexes, its validation chain and the identity of its sub-site.
Timestamp
A string giving the date and time a file was generated, embedded in its name following the convention (YYYY-MM-DD - HHhMM). Always in Luxembourg local time.
Token tokens
The base unit a LLM works with — roughly four characters, about three quarters of a word in English. French and German split less efficiently: a common word there costs closer to one and a half or two tokens. Context window, API billing and generation limits are all expressed in tokens.
Training
The phase in which a model is built, adjusting its internal parameters against vast text corpora. It happens once, up front, and stands opposite inference — the phase of use. A model learns nothing from your conversations: they do not alter its training.
Turn exchange
A unit made up of one user message and Claude's matching reply. A conversation is a sequence of turns.
Validator validators
A script that inspects a produced file and reports any anomaly. It returns a success or failure code, and a failure halts delivery.
Variability non-determinism
The property of a LLM whereby the same prompt can yield slightly different replies from one call to the next. Part of that is wanted: it allows creativity and avoids mechanical answers. Another part comes from the inference infrastructure itself and persists even at zero temperature.
WebView
The component that displays web pages inside a mobile application. Links within the family stay in it; external links are handed off to the phone's browser.
Working memory
The content active in the context window of a running session. This memory disappears when the session ends.
Working stretch working session
A unit of work inside a thread: a thread may carry several, over days or over a few hours. A stretch opens when work resumes, or when a project archive is deposited; the AI announces there the Kit version of its tree.
Working tree project tree
The set of the project's files the AI can work from: stylesheets, data modules, Registry, tools and produced documents. It is laid down once, as the project's archive, and stays available for as long as the working thread continues. That is where the source files live, and where every change departs from.
xAI
The American AI company founded in 2023 and the creator of Grok. xAI is tied to the X platform, whose stream its models draw on.