Claude Code Environment

A complete Claude Code environment: global guidelines, reusable skills, hooks, settings, version-controlled memory and learnings, and shared git configuration — everything symlinked into place from one repo.

Skills

---

Plan Implementation

Interactive design discussion that produces a structured plan document before any code is written.

Command: /plan-ralphex

Features:

• Researches the codebase to understand current architecture and patterns
• Asks clarifying questions and discusses design trade-offs
• Iterates on the approach until you're satisfied
• Outputs a plan in docs/plans/ with design decisions, file changes, and task breakdown

---

Create / Update PR

Prepares commits on a feature branch, pushes, and opens or updates a PR to main.

Command: /pr-create

Features:

• Reads the associated plan document and progress log for context
• Uses /reset and /commit to create clean, atomic commits from branch changes
• Rebases onto main before pushing so pr-merge can fast-forward
• Creates a new PR or updates the existing one (title, description, force-push)
• Drafts a detailed PR description with overview, changes, design decisions, and scope reconciliation

---

Merge PR

Merges a PR locally via fast-forward to preserve your GPG-signed commits.

Command: /pr-merge

Features:

• Avoids GitHub's merge UI, which re-signs commits with GitHub's own key
• Fast-forwards main to the PR branch (rebases as fallback if needed)
• Stashes uncommitted changes and restores them after merge
• Cleans up remote and local branches, prunes stale remote-tracking refs

---

Prepare PR

Summarizes what was done, how it matches the plan, and what the review found — all in one report before creating a PR.

Command: /pr-prepare

Features:

• Reads the plan doc, progress logs, and commits to build a complete picture
• Reports plan alignment: implemented items, unplanned additions, missing items
• Summarizes review findings: confirmed fixes, false positives, unaddressed concerns
• Read-only analysis — does not modify files or create commits

---

Clean Code

Audits modified files for dead code, duplication, and import hygiene.

Command: /clean-code

Features:

• Removes debug prints left from development
• Dead code audit: traces callers to find unreachable methods, fields, imports, type members
• Duplication audit: flags repeated logic and proposes consolidation
• Optimizes imports in modified files
• Iterates until clean — each removal pass may reveal new dead code

---

Commit

Analyzes changes and generates atomic Conventional Commit messages.

Command: /commit

Features:

• Reviews staged and unstaged changes, groups them into atomic commits
• Delegates to /reflect, /clean-code, and /documentation before planning commits
• Drafts commit messages in imperative mood with type prefixes
• Presents a full plan for approval before executing any commits
• GPG-signs all commits, never adds AI attribution

---

Update Documentation

Scans project documentation for stale references and fixes them.

Command: /documentation

Features:

• Checks README, docs/pages/, CLAUDE.md, and source comments against current code
• Fixes stale paths, API references, and behavior descriptions
• Keeps curated feature listings (features page, docs index, README) in sync with the diff
• Regenerates dimensioned-draft drawings when the model they document changed
• Suggests new documentation files or reorganization when beneficial

---

GitHub Pages Layout

Arranges a project's README and GitHub Pages docs into a consistent user-first layout — short README that links out, Jekyll site with a user-facing index, one page per user-facing feature, and exactly one developer page.

Command: /github-pages

Features:

• Enforces single-source-of-truth docs on GH Pages so the README stays under a screen
• Separates user-facing pages from the one developer entry-point page (development.md)
• Supports flat and monorepo variants with consistent navigation and screenshots
• Aligns new repos to the shape of existing reference implementations

---

Deploy

Configures and runs the local deploy — it makes the current code runnable on this machine and never ships anywhere public. Either installs the built app locally (Tauri, IntelliJ plugin, .NET) or starts the project's local web server. On first use in a project, sets up the deploy bash function, creates a local scripts/deploy.sh wrapper, and updates .gitignore.

Shipping outward is a different verb: a versioned artifact goes through the release skill, and a public site through the project's own CI — deliberately not a skill, since publishing from a working tree can leak uncommitted work or the wrong environment's credentials.

Command: /deploy

Features:

• Auto-configures deploy() shell function in the platform-appropriate rc file (~/.zshrc on macOS, ~/.bashrc on Windows Git Bash / Linux)
• Creates scripts/deploy.sh wrapper pointing to the global deploy script
• Reads install path from config/deploy.env (asks on first run)
• Install targets run the full pipeline: stop app → build → clean install dir → copy → launch → verify
• Local web servers (a package.json with a dev script, or a plain static index.html) relaunch detached on the configured port, then get health-checked
• After first /deploy, use ! deploy for instant deploys without LLM overhead

---

Cleanup

The destructive counterpart to deploy: stops the running app, removes the installed bundle, and wipes its user-data and cache directories — the "before a clean install" reset. Reuses config/deploy.env, so the install location is never entered twice.

Command: /cleanup

Features:

• Auto-configures cleanup() shell function in the platform-appropriate rc file (~/.zshrc on macOS, ~/.bashrc on Windows Git Bash / Linux)
• Creates scripts/cleanup.sh wrapper pointing to the global cleanup script
• Always asks which of bundle / app data / caches to remove — never assumes
• Backs up paths listed in BACKUP_FILES= to .cleanup-backups/<timestamp>/ before wiping data
• Recognizes Tauri projects; after first /cleanup, use ! cleanup for instant runs

---

Build

Configures a build shortcut for any project. On first use, sets up the build bash function, creates a local scripts/build.sh wrapper, and updates .gitignore. Then auto-detects the project type and builds. Optionally generates a GitHub Actions CI workflow.

Command: /build

Features:

• Auto-configures build() bash function in ~/.bashrc if missing
• Creates scripts/build.sh wrapper pointing to the global build script
• Auto-detects project type: npm, dotnet, or Tauri
• Optionally generates .github/workflows/build.yml with CI for push/PR builds (Tauri uses a Windows + macOS matrix)
• After first /build, use ! build for instant builds without LLM overhead

---

Release

Tags a new version, pushes to trigger CI, monitors the build, and updates the GitHub release with final notes. Supports dotnet and Tauri projects (Tauri builds for Windows + macOS).

Command: /release

Features:

• Validates preconditions: clean tree, on main, in sync with remote
• Auto-detects project type (dotnet or Tauri) and extracts project name
• Recommends version bump based on commit history, asks for confirmation
• Bumps version in all manifest files before tagging (csproj / package.json / tauri.conf.json / Cargo.toml)
• Creates signed annotated tags for GitHub "Verified" badge
• Compiles platform-appropriate release notes (SmartScreen + Gatekeeper first-launch warnings)
• Monitors CI (single-platform for dotnet, matrix for Tauri) until completion
• Replaces draft notes and un-drafts Tauri releases (GitHub auto-renders the assets list)

---

Document Data Flow

Generates or updates a data-flow architecture document (docs/data-flow.md).

Command: /document-data-flow

Features:

• Discovers the project's architecture by exploring the codebase
• Produces step-by-step flow diagrams with data transition annotations
• Generates message/API protocol tables for all message types and endpoints
• Follows strict formatting rules for consistency across updates

---

Reflect

Extracts durable knowledge from the current conversation and persists it to long-term memory before /clear or context compaction wipes it. Also runs automatically as an early step of /commit, so session learnings are captured alongside the changes they came from.

Command: /reflect

Features:

• Scans the conversation for feedback, project context, user profile, and external reference pointers
• Writes new memories or updates existing ones in global or project-scoped memory dirs
• Re-checks every project-scoped finding against "would this help in another repo tomorrow?", promoting what is actually global and splitting what is only partly so
• Audits already-stored project memories the same way, proposing promotions rather than performing them silently
• Flags candidate skill updates and learnings worth distilling
• Falls back to direct file reads when the gather-context helper is blocked by permissions

---

Memo

Parks an off-task idea in the project's memo backlog so it isn't lost — without derailing the current task — or lists the backlog to pick something up.

Command: /memo [idea]

Features:

• /memo <text> appends a dated - [ ] item to <repo>/.claude/memos.md (created on first use); /memo with no args lists the open backlog and offers to address one
• Memos are deliberately lighter than GitHub issues — half-formed thoughts, committed with the project
• Open items resurface on their own: at session start / /clear (via the memos-surface.py hook), at task completion, and after a /commit push
• Capturing a memo never starts the work — that's the point; addressing one is always an explicit, separate choice

---

GitHub Status

Cross-project overview of all your GitHub-owned local clones — branch, behind/ahead counts, uncommitted file/line totals, oldest pending work, and a per-repo description synthesized from the pending changes.

Command: /github-status

Features:

• Walks PROJECTS_ROOT (configured per-machine on first run), filters to repos owned by your GitHub user
• Fetches every repo's origin in parallel before reading state, so counts reflect the current remote
• Auto-pulls clean repos with inbound commits via git pull --ff-only, marks pulled repos with ✓
• Auto-hides columns that have no meaningful data (no unpushed commits → no UNPUSHED column, all on main → no BRANCH column, etc.)
• Reports uncommitted-file lists and unpushed-commit subjects so Claude can summarize each repo in one line

---

Update Plannotator Plugin

Force-updates the plannotator plugin by clearing stale caches and reinstalling.

Command: /plannotator-update

Features:

• Removes the marketplace cache (stale git clone that prevents updates)
• Removes the plugin cache
• Guides through reinstallation after restart

---

Transcrypt

Encrypts designated files with transcrypt so they are ciphertext in git history but plaintext in the working tree, or unlocks an already-encrypted repo after a fresh clone. See Encrypted memory for how this repo uses it.

Command: /transcrypt

Features:

• Uses one shared passphrase from Doppler (tools/prd → TRANSCRYPT_KEY), never a freshly generated one
• Marks files by the *.secret.* naming convention in .gitattributes
• Verifies the index holds ciphertext while the working tree stays readable, and stages without committing
• Relies on the global pre-commit guard as a safety net against committing an unencrypted secret

---

Hooks

External Hook Paths

When a hook command needs a path outside ~/.claude/ or this repo, reference it via a CLAUDE_<NAME> user-scope environment variable instead of hardcoding the absolute path. The hook command field is executed via shell, so standard $VAR expansion works — the same mechanism that already makes $HOME/.claude/hooks/... portable across machines.

Why: claude/settings.json is symlinked to ~/.claude/settings.json on every machine that uses this repo. Hardcoded absolute paths pin it to one user's filesystem layout; env vars keep it portable, and a repo move or rename only touches the env var (not every hook entry).

Caveat: This works for hook command strings only. It does not work for MCP server args in ~/.claude.json — those are passed straight to child_process.spawn() with no shell, so paths there must be absolute. That file is not symlinked from this repo.

Currently used env vars — set these on a fresh machine before the corresponding hooks will work:

• CLAUDE_AI_AGENT_DASHBOARD — points to a local clone of the tauri-dashboard repo. Used by the Notification, UserPromptSubmit, Stop, SessionEnd, and SessionStart hooks for live session-status updates.

Set on Windows (User scope, persistent):

[Environment]::SetEnvironmentVariable('CLAUDE_AI_AGENT_DASHBOARD', '{{path-to-tauri-dashboard}}', 'User')

Set on Linux / macOS (in your shell profile):

export CLAUDE_AI_AGENT_DASHBOARD="$HOME/projects/tauri-dashboard"

---

Memo Backlog

File: claude/hooks/memos-surface.py (one script, three modes)

Surfaces the open /memo backlog (.claude/memos.md, resolved at the git root, numbered newest-first) as a transient status-bar reminder so a fresh or freshly cleared session shows "what's next" without the user typing anything — then clears it the moment they start working. Three wired entry points:

• SessionStart (startup/clear, no arg) — writes a per-session state file with the open memos. Injects nothing into chat, so the model never greets with or pushes the backlog — the status bar is the only reminder.
• statusLine (statusline arg, refreshInterval: 2) — renders the compact backlog (top 3 + a +N more line) from the state file; the interval makes it appear within ~2s while the session is idle.
• UserPromptSubmit (on-prompt arg) — clears the state (the bar reminder is done). If the message is a bare number — or memo N / start N / do N / pick N — it injects, bound to that prompt, which memo N maps to, so Claude reliably starts it instead of treating the number as noise.

Stays silent when there's no file or nothing open. Needs no environment variable. See the Memo skill for how items get there and the other two moments they resurface (task completion, /commit).

---

Git Hooks

Pre-Push Validation

File: git/hooks/pre-push

Prevents pushing commits that are Claude-attributed or not GPG-signed. Every new commit in the push is checked for:

• Author or committer name/email containing "claude" or "anthropic"
• Co-Authored-By trailers mentioning Claude or Anthropic
• Missing good GPG signature (only G status passes)

Global installation is covered in the Global Installation section below.

---

Encryption guard (pre-commit)

File: git/hooks/pre-commit

Installed by transcrypt and made portable: in a repo configured for encrypted memory it blocks a commit if a *.secret.md file is staged unencrypted (the last guard against a plaintext leak). Because core.hooksPath is global, it guards on the per-repo transcrypt copy and is a no-op in any repo that doesn't use transcrypt.

---

Learnings

The claude/learnings/ directory collects long-form, domain-specific reference notes — non-obvious behaviors learned through trial and error (framework quirks, API limitations, platform gotchas), each a topic-named markdown file with no frontmatter or index. They're available globally through the ~/.claude/learnings/ symlink, and the /reflect skill adds to them as new knowledge surfaces.

The filenames are the index — browse claude/learnings/ to see what's covered rather than maintaining a manifest here. To pull a topic into a project, point that project's CLAUDE.md at the file:

Read `~/.claude/learnings/chrome-extension.md` for domain-specific patterns.

---

Global Installation

Global files live in claude/ (symlinked to ~/.claude/) and git/ (hooks, gitignore, gitattributes — each symlinked to ~/). Project-local config stays in .claude/.

If any of these already exist in ~/.claude/ or ~/.git-hooks/, move them into the repo first (or remove them) before creating the symlink.

macOS / Linux users skip this section — see Linux / macOS below.

Windows

Run from the project root as Administrator:

New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\CLAUDE.md" -Target "$PWD\claude\CLAUDE.md"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\skills" -Target "$PWD\claude\skills"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\hooks" -Target "$PWD\claude\hooks"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\settings.json" -Target "$PWD\claude\settings.json"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\learnings" -Target "$PWD\claude\learnings"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\memory" -Target "$PWD\claude\memory"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\scripts" -Target "$PWD\claude\scripts"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.git-hooks" -Target "$PWD\git\hooks"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.gitignore" -Target "$PWD\git\gitignore"
New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.gitattributes" -Target "$PWD\git\gitattributes"
git config --global core.hooksPath "$env:USERPROFILE\.git-hooks"
git config --global core.excludesFile "~/.gitignore"
git config --global core.attributesFile "~/.gitattributes"

Linux / macOS

Run from the project root:

mkdir -p ~/.claude
ln -s "$(pwd)/claude/CLAUDE.md" ~/.claude/CLAUDE.md
ln -s "$(pwd)/claude/skills" ~/.claude/skills
ln -s "$(pwd)/claude/hooks" ~/.claude/hooks
ln -s "$(pwd)/claude/settings.json" ~/.claude/settings.json
ln -s "$(pwd)/claude/learnings" ~/.claude/learnings
ln -s "$(pwd)/claude/memory" ~/.claude/memory
ln -s "$(pwd)/claude/scripts" ~/.claude/scripts
ln -s "$(pwd)/git/hooks" ~/.git-hooks
ln -s "$(pwd)/git/gitignore" ~/.gitignore
ln -s "$(pwd)/git/gitattributes" ~/.gitattributes
git config --global core.hooksPath ~/.git-hooks
git config --global core.excludesFile "~/.gitignore"
git config --global core.attributesFile "~/.gitattributes"

Memory

Two complementary stores hold accumulated cross-session knowledge, both surfaced to the harness for auto-recall:

• Global memory (claude/memory/, deployed to ~/.claude/memory/ via symlink) — cross-project preferences, feedback, and references meant to apply everywhere.
• Project memory — facts specific to a single repo. Claude Code writes these to a machine-local cache (~/.claude/projects/<path-encoded>/memory/) that is not version controlled, so the knowledge is invisible from other machines and lost if the cache is cleared.

Versioning project memory

The claude/scripts/link-project-memory.sh script redirects a project's memory cache — via a symlink, or a directory junction on Windows — into a committed .claude/memory/ directory inside that repo. The harness keeps reading and writing the same path, so auto-recall is unaffected; the files just live in the repo now and travel with git clone.

Run once per project, per machine — from inside the repo:

bash ~/.claude/scripts/link-project-memory.sh

It migrates any files already in the cache, wires up the link, and leaves .claude/memory/ staged for you to commit. On a fresh machine, clone the repo and re-run the command to re-establish the (machine-local) link.

In this dotfiles repo the two stores sit side by side: claude/memory/ is the global payload deployed to ~/.claude/memory; the repo-root .claude/memory/ is this repo's own project-specific memory.

Encrypted memory (*.secret.md)

Memory files holding sensitive coordinates (not secret values — those stay in Doppler) are committed encrypted, so this public repo never exposes them. They are transparently decrypted in a working tree that holds the key, and read as opaque blobs to anyone without it.

• Mechanism: transcrypt (vendored at claude/scripts/transcrypt) wires Git clean/smudge filters. .gitattributes marks claude/memory/*.secret.md filter=crypt, so those files are ciphertext in every commit and plaintext only locally.
• Key: a symmetric passphrase kept in Doppler (a TRANSCRYPT_KEY secret) — not in this repo. The committed index entry for an encrypted memo is deliberately generic, so even the description gives nothing away.

Unlock on a new machine — after cloning and fetching the key, run from the repo root (substitute your Doppler project/config):

bash claude/scripts/transcrypt --yes -c aes-256-cbc \
  -p "$(doppler secrets get TRANSCRYPT_KEY --project <project> --config <config> --plain)"

The portable pre-commit (installed globally via core.hooksPath) already chains transcrypt pre_commit, so init skips writing its own redundant helper hook into the shared, committed git/hooks/ — no leftover, no manual cleanup. (This is a local patch to the vendored transcrypt; re-apply it if you re-vendor upstream.)

Until then, *.secret.md files read as encrypted blobs. Add more by naming them *.secret.md; the attribute pattern encrypts them automatically.

License

GPL-3.0