原始内容
Talent Tree
Overpowered Claude Code talents to power level your software architecture.
Install Talents
Requires Claude Code.
/plugin marketplace add heyJordanParker/dotfiles
/plugin install talents@talent-tree
Talents
Invoked as /talents:<name>:
- agent-browser — Automate browser interactions, web testing, screenshots, data extraction (requires agent-browser)
- architecture — Present architectural options with tradeoffs
- cc — Work with Claude Code skills, hooks, settings
- code-first — Autonomous async execution mode for AFK / overnight handoffs
- codex — Drive codex CLI runs as your agents
- commit — Stage, write the message, commit, verify; structured commits with type prefix and file tree
- debate — N independent architects debate architectural options through structured rounds with cross-pollination
- debug — Systematic debugging with root cause tracing
- design — UI components, styling, and interaction decisions
- diagram — Generate Excalidraw diagrams on a live canvas (requires drawbridge)
- execute — Contract for executing-state turns: implement approved work, escalate on architectural change
- execute-plan — Orchestrate implementation plans with persistent team, validation gates per slice
- gh-issue — Create self-contained GitHub issues an agent can implement without conversation history
- impeccable — Design, audit, and polish frontend interfaces
- independent-review — N identical parallel agents for consensus through redundancy
- laravel — Laravel PHP patterns for controllers, models, migrations, queries
- modeling — Transform shaped parts into concrete models (DB schema, UX flows, architecture)
- naming — Consistent naming for variables, files, classes, everything
- pcc — Add pros/cons/confidence to any prompt
- personas — 5 parallel persona agents for diverse perspectives
- pragmatic-engineering — KISS-driven planning and review
- propose — Contract for proposing-state turns: the seven named proposal failures and the choice-block shape
- react — React patterns for components, hooks, data fetching, state, performance
- regressions — Map a diff to affected flows and report capability regressions
- review — Parallel code reviewers on uncommitted changes
- review-plan — Review planning artifacts with 5 parallel specialized agents
- shaping — Collaborate on problem definition and solution options
- show-architecture — Annotated file trees inline
- slicing — Break modeled features into vertical implementation slices with acceptance criteria
- solo — Single-agent deep-research mode, no subagents
- subagents — Framework for dispatching one-shot subagents
- trace — Code intelligence CLI: search, callers, definitions, complexity, file/method reads with rich architectural context (requires the bundled
tracebinary; see Tracer Setup below) - user-testing — Trace real user flows through code changes, find gaps
Commands
Also invoked as /talents:<name>:
- ask — Break complex scenarios into decision questions with 4+ options
- commit — Validated commit with tests and comprehensive review
- copy-plan-path — Copy the current plan path to the clipboard
- copy-shaping-dir — Copy the current shaping directory path to the clipboard
- copy-transcript-path — Copy the current conversation transcript path to the clipboard
- plan — Plan a feature using structured format
- retro — Analyze conversation history for patterns and improvements
- wtf — Hard reset, redo the last task correctly
Hooks
The plugin includes hooks that run automatically to keep Claude disciplined:
- block-git-revert — Blocks
git reset,git restore,git checkout -- <file>. Forces manual execution. - block-unsafe-delete — Whitelists
rmto specific directories only. Everything else blocked. - validate-planning-docs — Checks planning documents on write/edit for deferral, optionality, and other quality issues.
- validate-plan-quality — Gates
ExitPlanModeon plan quality before a plan is accepted. - sync-shaping — Ripple-check reminders when editing shaping documents
Tracer Setup
The /trace skill needs the trace binary. The plugin ships a POSIX launcher that lands on PATH automatically when the plugin is enabled. tracer is a native Rust binary, so the launcher resolves a host-appropriate one: a committed prebuilt (macOS arm64 / Linux x86_64), a previously locally-built cached binary, or — on any other platform — cargo build from the crate source shipped in the plugin. If no Rust toolchain is present for that last path, trace prints an actionable "install Rust" error.
All 23 commands work from whichever binary the launcher resolves — there is no reduced-feature tier. If you want trace outside the plugin too, build it from tools/tracer and put it on PATH (the dotfiles setup.sh does this into ~/.local/bin):
cd tools/tracer
cargo build --release
install -m 755 target/release/trace ~/.local/bin/trace
Tracer wraps five external binaries: ast-grep, scc, universal-ctags, ripgrep, git. Install whichever your platform needs:
macOS (Homebrew):
brew install ast-grep scc universal-ctags ripgrep
xcode-select --install # git
Linux (Debian/Ubuntu):
apt install universal-ctags ripgrep git
# ast-grep: https://ast-grep.github.io/guide/quick-start.html
# scc: https://github.com/boyter/scc#installation
Windows (Scoop):
scoop install ast-grep scc ripgrep
# universal-ctags: https://github.com/universal-ctags/ctags
# git: https://git-scm.com/download/win
Verify with trace doctor — it lists missing binaries with the same install command for your platform.
Agents
Per Claude Code plugin schema, agents (.md files in packages/agents/agents/) are NOT distributed by the marketplace install — only skills, commands, and hooks. To use the bundled agents (explorer, researcher, architect, debugger, etc.), copy the files you want into your own ~/.claude/agents/ or <repo>/.claude/agents/:
git clone https://github.com/heyJordanParker/dotfiles.git /tmp/dotfiles
cp /tmp/dotfiles/packages/agents/agents/explorer.md ~/.claude/agents/
cp /tmp/dotfiles/packages/agents/agents/researcher.md ~/.claude/agents/
# ... and any others you want
The split worth knowing:
- explorer — in-codebase architectural mapping ("where is X used", "how does Y work end-to-end"). Uses the trace skill heavily.
- researcher — external research (library docs, APIs, framework references, web lookups). Uses agent-browser, cc, claude-api, plus trace for incidental in-repo grounding.
- architect / backend-engineer / code-reviewer / debugger / frontend-engineer / regression-reviewer / tester — also have the trace skill for navigating the codebase during their work.
Safe Delete
Claude loves running rm -rf. Protect yourself by replacing rm with trash so deleted files go to macOS Trash instead of being permanently destroyed:
brew install trash
Then alias rm to trash in your shell profile:
cat >> ~/.zshenv << 'EOF'
# Safe delete - moves to Trash instead of permanent deletion
rm() {
local args=()
for arg in "$@"; do
[[ "$arg" =~ ^-[rRfidv]+$ ]] && continue
args+=("$arg")
done
trash "${args[@]}"
}
EOF
This silently strips rm flags (-rf, -i, etc.) and sends files to Trash. Claude thinks it's deleting; you can undo from Trash.
Dotfiles
macOS dotfiles managed with GNU Stow.
New Machine Setup
One-line bootstrap (clones to ~/dotfiles, then runs setup):
curl -fsSL https://raw.githubusercontent.com/heyJordanParker/dotfiles/master/setup.sh | bash
To clone elsewhere, do it yourself first; setup will run from whichever path you pick:
git clone https://github.com/heyJordanParker/dotfiles.git <path>
cd <path> && ./setup.sh
Either way installs Xcode CLI tools, Homebrew, all packages from Brewfile, bun, and symlinks configs.
Manual Usage
Each subdirectory of packages/ is a stow package. Its contents are exactly what should land at the package's stow target — no inner mirror wrappers. The package-to-target mapping lives in scripts/stow.py.
# Add a new config — example: 'newapp' targets ~/.config/newapp/
cd <repo>/packages
mkdir -p newapp
mv ~/.config/newapp/* newapp/
stow -t ~/.config/newapp newapp
# Remove symlinks
stow -D -t ~/.config/newapp newapp
# Re-link after changes
stow -R -t ~/.config/newapp newapp
Structure
Stow packages live under dotfiles/packages/<pkg>/. Each package contains the actual files that get symlinked to its target — see Claude.md for the full architecture.