Add-ons
Opt-in extras. Nothing here is installed until you ask.
Chapter 1
Off until you switch it on
Each add-on is installed on request, never as part of the default kit.
Why they are opt-in
Every add-on below is marked opt-in only in its own file. Some add a hook that runs in every session, some download a model, one needs an API key or a local server. None of that should happen unasked.
Once you turn one on, it behaves like any installed skill and stays until you remove it.
Add-on 1 of 9
Personality
Replies phrased to fit the reader.
Does
Six original personas change how Claude words its replies: two voices (female, male) times three skill levels (novice, mid, pro). Wording only: facts, code, commands, paths, numbers, error text and how carefully work is verified never change.
You
Pick one of six personas, none, or a custom one at install. Run node setup.js --install <choice> --hook.
AI
A SessionStart hook adds the chosen persona, about 150 to 190 tokens, to every session. Nothing is added until a persona is chosen.
Levels and custom personas
novice: plain words, terms explained once, reason before step. mid: everyday dev terms unexplained, reason and fix together. pro: full jargon, result first, exact names, no preamble or recap.
An install copies one persona, because a session needs one. Edit ~/.claude/personas/custom.md for your own, or put a file named like female-pro.md in ~/.claude/personas/ to override a built-in. Keep each under about 100 words, since the text is added to every session.
The text is written to read well aloud, so it pairs with narrate: pick the same gender for the voice.
Add-on 2 of 9
Narrate
Replies read aloud.
Does
A Stop hook reads the final reply aloud. Code, tables and links are dropped and announced as on screen. It runs locally, never talks to Claude, and costs no tokens.
You
Run node setup.js --voices (about 125 MB of free, offline Piper voices), then node setup.js --hook. Control it with /narration: mute, voice, volume, speed, pitch, play, pause, resume, mode and alerts.
AI
Nothing. The default mode is manual: the hook only keeps the reply until you play it. Mode auto speaks every reply as it arrives.
Platforms and your own voice
Playback follows the system default output on Windows, macOS, Linux and Android (Termux). Only Windows has been run on a real machine; the other rows are covered by tests of the player choice, not by listening.
You can add another Piper voice, or a cloned voice through a command engine that reads text on stdin and writes a wav. Licences are yours to check; never commit a custom voice or its reference audio to a shared repo.
pause holds the speech and resume continues from the same spot; stop still cuts it and clears the queue.
Spoken alerts are off by default. When a reply ends on a question, a short line in a second voice says so, only after the keyboard has been idle (default 30 seconds, Windows). An optional alertCommand sends alerts, and only alerts, to your own engine; every rendered piece is cached under ~/.claude/narrate-cache.
For on-screen buttons in VS Code, install extensions/narrate-buttons.
Add-on 3 of 9
Project context
Private project docs, indexed for few tokens.
Does
Keeps requirements, brand notes, assets and Figma exports in one gitignored project-context/ folder. Each sub-folder gets an INDEX.md, so the agent finds a fact without reading the raw files.
You
Ask for the feature, choose auto-detect, recommended or your own folders, drop docs in, then run /ctx analyze.
AI
Reads project-context/INDEX.md at session start, answers from each folder's Key facts, and edits INDEX.md files only, never your docs.
The tool and the privacy rules
scan.js does the mechanical work: detect, init, status, changed, stamp, root and add. The agent reads only files that are new or modified since the last analysis.
The folder is gitignored and pre-commit blocks any staged file under it. Its content never goes into a commit, PR, published page or subagent prompt unless you say so. If a file holds credentials, the agent says so without printing them.
Where two docs disagree, both are listed under conflicts and the agent does not pick one.
Add-on 4 of 9
System-1 prefilter
Rank candidates with a typed-decision model.
Does
Takes a state and typed questions and returns a chosen option with a probability per option, never generated text. It cuts file reads before they cost tokens.
You
Choose a hosted endpoint with an API key, or local Laya: free, offline, no key, about 1 GB of weights on CPU.
AI
If unconfigured, it skips prefiltering and does the task the normal way. A deterministic ladder decides when to read the runner-up.
What it is not for
Proven uses: narrowing files after a search, code-review triage, test selection, skill-applicability decisions, and binary or multiple-choice questions already asked in that shape.
Never for security or safety pass-fail checks: a real miss was found there, a plaintext secret scored a wrong pass. It is also not for choosing between solution approaches.
It must never be pointed at a general chat model, because those do not return calibrated probabilities.
Add-on 5 of 9
Diagram
Stack and structure diagrams drawn by a script.
Does
diagram.cli.js detects a project's stack or folder structure into a spec, lays it out, checks it for overlaps, arrows through text and a missing accessible title and description, and renders light and dark SVG plus a self-contained HTML page and a README snippet.
You
Type /diagram detect, confirm the guessed links, then /diagram build <spec>. Logos are fetched only if you agree to /diagram icons.
AI
Supplies roles, wording and confirmation of guessed edges, and looks at the result before calling it done. It never hand-writes SVG.
Why a script draws it
Hand-drawn diagrams put arrows through text, drift from the code and cost tokens on every redraw. A script lays out and checks the same spec every time.
Builds make no network calls. Brand logos are cached once with the icons word.
Add-on 6 of 9
Figma Code
An exported Figma SVG turned into pixel-matched HTML and CSS by a script.
Does
figma-code.cli.js reads the SVG, writes index.html and styles.css, renders both in headless Chrome or Edge and diffs them pixel for pixel. PASS means the page matches.
You
Export the frame as SVG with Outline text unchecked, then type /figma-code build <path.svg>, adding --download-fonts to fetch Google Fonts. Give the path, never an attachment.
AI
Runs the script and reports what the diff says. It does not draw or tune layout.
Fonts and project-context
An SVG holds font names, not files. A missing font looks like a layout bug in the diff, so the script lists missing fonts and fetches them only with --download-fonts.
With project-context, the SVG lives in project-context/design/ and each build records one row per frame in figma-code.md.
Add-on 7 of 9
Modern Web
Finds legacy web patterns and lists what the platform now does natively.
Does
modern-web.cli.js runs Chrome's pinned modern-web-guidance scanner over your folder, drops vendor and minified files, caches the result for 24 hours and prints a short table. Every row is a candidate, not a verdict.
You
Type /modern-web, or run node cli.js with --changed, --limit, --refresh, --json or --dry-run. Needs Node 22 or newer. The first run downloads the scanner, about 9 MB.
AI
Relays the table and checks each row's keepWhen note before touching code. It edits only when you ask to modernize. Where the kit's standard disagrees, the kit wins.
Limits
The scanner covers about 20% of the guidance, so no findings does not mean the code is modern.
Offline, or with an old Node, the scan is skipped with one line saying why. The scanner is by the Chrome team, is called at run time and is never copied.
Add-on 8 of 9
Kaggle
Claude plans and reviews; a free Kaggle GPU model writes the files.
Does
Runs a plan's coding tasks in a private Kaggle notebook with qwen3-coder:30b on two T4 GPUs. Files land in a staging folder, never the repo, and local gates check placeholders, syntax and missing files. Failed tasks retry twice, then Claude takes over.
You
Make a free Kaggle account with a verified phone, save the API token in .env, then /kaggle setup, /kaggle plan plan.md (free) and /kaggle run plan.md. You start every step.
AI
Writes the plan in the task format, reviews every staged file, fixes what is left and copies files into the repo only after checking them.
Plan format and quota
One section per task with Target, Check, Standards and Context lines. Each task is one or two files and stands alone, because the model cannot ask questions.
GPU quota is about 30 hours a week free. run refuses above 90% use or while a session is open unless forced. It lives under the kit's cloud/ folder, where features that call an internet service go.
Add-on 9 of 9
Teardown
Interactive 3D simulators of any machine or model, built on Triforge.
Does
teardown.cli.js detects the project, scaffolds one machine-agnostic page where any set of named parts gets isolate, a cutaway, a camera framed from the real bounds and a tour stop, then checks, builds an offline page, and renders review sheets in headless Chrome.
You
Describe the machine, or give a .glb or .gltf model. Run /teardown detect, init, check, build, sheet and probe. Starters: gears, crank or custom.
AI
Supplies the real dimensions, the reference and the machine's geometry, and reads the sheet before calling it done. Any probe FAIL is a bug in the page.
Film and sound
teardown-film records the tour to an H.264 MP4 by stepping the page's pure render(t) in headless Chrome and encoding with ffmpeg.
teardown-sound writes a reproducible score WAV from the tour's timeline, mastered to -14 LUFS with the peak at or below -2 dBFS. Neither is part of a default install, and both need teardown.
Finale
Not sure which one?
Start from the situation.
You want replies worded for your skill level
personalityYou want replies read aloud
narrateThe project comes with requirements, brand and Figma notes
project-contextYou want fewer file reads before work starts
system1-prefilterYou want an architecture or structure diagram for a README
diagramYou want an interactive 3D cutaway of a machine or model
teardown
Keep reading
On to the next page
The docs are a short run of pages. Use the buttons, or jump to any page.