Moddeck

Safety and productivity mods for Claude Code, built for scientists. Thirty-seven small add-ons (mods), five skills and a set of git guards that protect your work, keep your token bill visible, and stay quiet until they have something useful to say.

Credit where it is due. Not everything here is my own creation. Some mods and skills are adapted from other people's ideas and guides, and some features already exist inside Claude Code or as Anthropic's official plugins: Moddeck simply switches them on and makes them easier to reach. skins (fifteen colour themes, picked from a dropdown) is hellosverre/claude-skins (MIT, © 2026 hellosverre), copied with its update check removed. Seven mods (staged-diff, contrast-check, table-preview, test-ledger, scope-watch, error-inbox, decision-log) are adapted from whyashthakker/awesome-claude-code-mods (MIT, © 2026 Yash Thakker and contributors), rewritten for moddeck; each keeps the original licence and says what changed in its README.

Install

1. Make sure you have pipx, git, Python 3.9+ and Claude Code 2.1.287 or newer.

2. Run:

   pipx run --spec git+https://github.com/samuelbfernandes/moddeck@main moddeck install --standard

3. Start a new Claude Code session.

• Look before installing: add --plan (prints every change, then stops).
• Pick items yourself: drop --standard and answer a few questions.
• Update: click Update now when moddeck offers it, or run the same line with update --scope user -y instead of install --standard.
• Remove: run the same line with uninstall; it removes exactly what it installed.
• Other ways (uv, pip, conda, a clone, only the mods through /plugin): docs/INSTALL.md.

The side bar

/moddeck opens one side bar for the whole toolkit:

• Session: model, effort, permission mode, context, plan usage with reset times, cost today and this week, cache, git branch, backup age, tokens, recent tools and your environment.
• Side Question: ask something without adding it to the chat.
• Mods, in grouped cards. A name is its switch: green is on, red is off, one click flips it. A gray name is a command that runs when clicked.
• Skills and saved prompts: click a /name to put it in the prompt box.

In the Claude Code desktop app

1. Start a session: the side bar opens by itself beside the chat.
2. Type /moddeck to close or reopen it.
3. Using the app in light mode? Click ☀ at the top of the side bar (☾ for dark), or type /moddeck theme light.
4. Drag its edge to resize: text that does not fit ends in "…".

In the terminal

1. Type /tui fullscreen.
2. Type /moddeck: the side bar docks on the right.

What is included

Commands always start with /; a word without the slash is an ordinary prompt. Turn any mod off with /plugin disable <name>. Every tool has a page in docs/TOOLS.md.

Mods

Prompt helpers

• model-picker: suggests a cheaper or stronger model when your prompt calls for it, and flags an effort level that does not fit.
• reflect: offers to save a rule when you correct Claude (spends a few tokens).
• tldr: a dropdown, off until you pick a mode. highlight draws each long reply's must-read sentences bold with ⚑; summary also adds a TL;DR box above the prompt (Summary, Needs you, Next). One Haiku call per long reply.
• skins: a dropdown of fifteen colour themes (noir, tokyo-night, dracula, nord, and more) that also redraw tool rows, diffs, tables and code; off returns Claude Code's own look. By hellosverre, MIT.
• concise: a switch, off until you turn it on; then Claude's replies lead with the next action, number the steps, restate where you are and leave tangents for later (adapted from ayghri/i-have-adhd, MIT).
• skill-router: a prompt that clearly asks for a toolkit skill (review, plan-first, writing, ahpcc, assess-plugin, agents-md-scaffold) gets one line naming it; local keywords, once per skill per session.
• prompt-upgrade: /better <prompt> turns a rough prompt into a great one: up to three multiple-choice questions, a rewrite, then your prompt and the upgrade side by side in a pane. Nothing runs until you press Run. /better on sends every typed prompt there first.
• prompt-library: /prompts keeps one list of saved prompts for every project, one click away in the side bar.

Model controls

• command-guard: before a risky command, shows exactly what would be lost, with Cancel first.
• review-cost: shows the token cost of a Codex review before a push.

Guards

• secret-guard: hides keys, tokens and passwords from what Claude reads.
• method-review: when Claude edits method code (paths with blup, herit, selection, variance and similar words), the first git commit or git push is held once until a second model has reviewed it.
• verify-before-done: a toast, and a reminder to Claude on your next prompt, when a turn changed code and ran no test, check or build.
• scope-watch*: /scope-watch <folder>; a note when Claude edits a file outside it.
• data-guard: asks before a commit holds large files, data files, personal-looking column names or notebook outputs.
• big-file-guard: reads only the first lines of a big csv, json or log.
• repro-log: /repro writes a re-runnable recipe of the session (commands, versions, seeds).

Automation

• auto-test: runs your tests after a turn that changed code (off until you switch it on).
• long-run-ping: a toast, sound and notification when a long run finishes.
• format-on-edit: formats each file Claude edits with Air (R), Ruff (Python), or the project's Prettier or Biome (off until you switch it on).
• test-ledger*: every test, lint and build run of the session with its result (R, Python, Rust and JS tools).

Workspace utilities

• checkpoint: /checkpoint snapshots your working tree; a restore shows what would change first.
• branch-janitor: notices a finished branch, checks it is safe, offers to delete it.
• cache-guard: warns before a message that would re-pay for a cold prompt cache.
• usage-meter: a quiet usage line and an optional spending limit.
• progress: a plan progress bar and a tree of helper agents with estimated cost.
• filetree: a live folder tree beside the chat.
• edit-replay: replays Claude's edits one diff at a time.
• diagram: draws Claude's mermaid diagrams as box art, offline.
• staged-diff*: the staged diff, before you commit.
• error-inbox*: the last 30 failed or refused tool calls, secrets scrubbed, with an "ask Claude" button.
• decision-log*: /decision <text> adds a dated row to docs/DECISIONS.md.
• table-preview*: /table-preview data.csv previews a table: delimiter, column types, first rows.

UI development (more in docs/UI-DEV.md)

• ui-preview: runs the app in this project (plumber API first, then a JS dev server, Shiny, Streamlit), lists the plumber routes, takes a screenshot and asks Claude to review the page.
• contrast-check*: colour formats and WCAG contrast, with the nearest passing shade.

Look and feel

• moddeck: the side bar above.

* adapted from awesome-claude-code-mods (see the credit note above).

Skills (Claude loads them by itself)

• review: a quick self-review for ordinary code; a second model (Codex or a fresh read-only Claude) for genetics, statistics and method changes.
• plan-first: interviews you before a big build, helps choose the design, then writes a spec and hands off to a fresh session.
• writing: options (three graded versions), edit and detect (takes the AI tells out, keeps your voice; includes the humanizer patterns), and rewrite, in one skill.
• agents-md-scaffold: sets up AGENTS.md, a decisions log and a review rubric.
• assess-plugin: what a plugin installs, costs and risks, before you install it.

Git guards and settings

• Git guards: refuse a committed .env or private key, and force-pushes, in every repository (or just one).
• Push checks: git push asks first; R and Rust projects run their tests before a push.
• Hardened settings: Claude's file tools cannot read SSH, cloud, GPG or .env files.
• Claude Code's own features, switched on for you if you tick them: You Should Know tips, the sandbox, telemetry off.
• Anthropic's official plugins (optional): Rust and Python diagnostics, security review, CLAUDE.md upkeep, code simplifier, PubMed, bioRxiv, Context7.
• UI development extras (optional, third-party, never on by default): Chrome DevTools MCP (pinned, usage statistics off), Posit's btw and mcptools for R, Posit's Shiny, Quarto and R-package skills, Plotly Cloud MCP, Anthropic's frontend-design. Pick them with --with <id>; the list and the programs to install yourself (Air, Ruff, jarl, the Shiny for Python, Streamlit, Gradio and Panel skills) are in docs/UI-DEV.md.

Try it

Work in a scratch repository so nothing real is at risk:

mkdir demo && cd demo && git init && claude

Then try any of these:

• moddeck: type /moddeck and click the green suggest name. It turns red (off) at once.
• command-guard: ask create a folder build with a file a.txt, then delete the build folder. A box shows the risk and what would be lost, with Cancel first.
• model-picker: on Opus, type fix the typo in the README title. A dialog says Opus is overkill and suggests a cheaper model.
• secret-guard: type my token is [secret:github-token], remember it. A dialog offers to send it redacted.
• data-guard: run echo "email,age" > people.csv && git add people.csv, then ask Claude to commit. A dialog names the file and the column email.
• checkpoint: type /checkpoint, edit a file, then /checkpoint restore <name>. It lists the files that would change.
• branch-janitor: merge a finished branch into main, then send any prompt. It offers to delete the branch.
• usage-meter: type /usage-meter 80, then /usage-meter for the report.
• diagram: ask for a mermaid flowchart of a GBLUP analysis, then type /diagram.
• Side Question: type /btw what is REML?. The answer appears in the side bar, not in the chat.
• Git guards: run echo KEY=1 > .env && git add .env && git commit -m test. The commit is refused.

Each mod's own README (mods/<name>/README.md) has its settings and limits.

Why you can trust the installer

• It shows the full plan first (--plan) and asks before applying.
• No network, no sudo, no third-party code unless you tick it in the UI development list: standard-library Python, checked by CI.
• uninstall removes exactly what it recorded and leaves your own settings alone.
• docs/AUDIT.md lists, for every mod, the programs it starts and whether it writes files or spends tokens.
• docs/SAFETY.md lists what it touches, what it never does, and the real caveats.

Popular, but think twice

The installer never installs these. Each has a cost or catch worth reading first (sources and dates: docs/SUGGESTIONS.md).

• obra/superpowers: skills that trigger on their own; a bootstrap at every session start with no published token figure.
• thedotmack/claude-mem: automatic memory; runs model summaries of your tool use and stores them.
• github/spec-kit, Fission-AI/OpenSpec: spec-driven workflows; many manual steps, and OpenSpec has telemetry on by default.
• ruvnet/ruflo, SuperClaude Framework: agent swarms; large always-present instructions and background workers.
• musistudio/claude-code-router: routes requests across providers; built for API keys, and model-picker does the cheap version.
• openai/codex-plugin-cc: Codex commands; its review gate reviews every reply and can drain usage limits.
• wshobson/agents, davila7/claude-code-templates: catalogs of 100+ agents and skills; browse, do not bulk-install.

Worth a look if you want them: trailofbits/claude-code-config, kenryu42/cc-safety-net, K-Dense-AI/claude-scientific-skills (install a subset), ccusage and claude-hud / ccstatusline. posit-dev/skills is now an optional installer item (see above).