Claude Code Workshop — Starter Kit

Everything you need to replicate the workflow from the workshop. Run the installer and you're up in under two minutes.

bash install.sh

---

Files

install.sh

One-shot installer. Run it once on a new machine and it handles everything:

• Installs Claude Code via the official native installer
• Installs haft (decision engineering MCP server)
• Installs uv (Python package manager, required by the Beads plugin)
• Installs Beads (issue tracker)
• Copies CLAUDE.md to ~/.claude/CLAUDE.md (backs up any existing file first)
• Copies the anti-deskilling skill to ~/.claude/skills/anti-deskilling/ and runs its hook installer (requires jq — install with brew install jq first)
• Tries to install the Inter font via Homebrew if available

---

CLAUDE.md

The global agent instructions file — install to ~/.claude/CLAUDE.md and it applies to every project on your machine automatically. This is Elias's actual working file.

It covers:

• Anti-Deskilling Gate: forces Claude to invoke /anti-deskilling before architecture, algorithm, security, or irreversible decisions. Also hooked on every Write/Edit.
• Communication style: peer engineer mode — direct, dry, no cheerleading, no hedge phrases. Includes a calibration table of preferred vs. avoided language.
• Thinking principles: Separation of Concerns, Weakest Link Analysis, Explicit Over Hidden, Reversibility Check — applied to every non-trivial problem.
• Task execution workflow: 8-step protocol from understanding the problem through investigation, planning, implementation, debugging, and verification.
• haft (FPF) reference: full command table, recommended cycle, inline quick-decision framework, and key concepts (R_eff, evidence decay, CL penalties, Transformer Mandate).
• Critical reminders: the 9 rules that matter most, always visible at the bottom of context.

Note: Beads auto-generates a project-level CLAUDE.md with issue tracking instructions when you run bd init. You don't need to add that manually.

---

user-settings.json

Template for ~/.claude/settings.json — your personal global defaults that apply to all projects.

Copy to ~/.claude/settings.json and adapt:

• Sets default model to Sonnet and effort to medium
• Pre-approves safe read-only tools and common build commands (git *, npm test, pytest)
• Sets DISABLE_TELEMETRY=1 and DISABLE_ERROR_REPORTING=1 via the env key

This is the minimum viable config. Add more allowed tools as you learn what you trust.

---

project-settings.json

Template for .claude/settings.json inside a project — shared with your team via git.

Copy to .claude/settings.json in your repo root and adapt:

• Whitelists project-specific safe tools (make *, cmake *, ctest *)
• Includes a SessionStart hook that runs bd prime to load open Beads issues into context at the start of every session
• This file is committed to version control so the whole team gets the same tool permissions

Unlike user-settings.json, this is per-project and should reflect what's actually safe in that codebase.

---

anti-deskilling/

Elias's actual anti-deskilling skill — a complete, working example of a skill with hooks, evals, and a reference guide. The installer copies this to ~/.claude/skills/anti-deskilling/ and wires the hooks globally.

What it does:

• PreToolUse hook — intercepts every Write and Edit call and checks whether Claude is about to implement something you should be reasoning through yourself (architecture choices, algorithm selection, security patterns, irreversible decisions). If so, it gates and asks.
• UserPromptSubmit hook — resets the engagement counter at the start of each prompt so the gate doesn't accumulate across unrelated tasks.
• /anti-deskilling — invoke manually at any time to trigger the engagement check explicitly.

Structure inside the folder:

• SKILL.md — the skill instructions Claude reads
• scripts/install.sh — hook installer (run with --global for user-wide, default is project-level)
• scripts/uninstall.sh — cleanly removes hooks from settings.json
• scripts/engagement-gate.sh — the PreToolUse hook script
• scripts/reset-counter.sh — the UserPromptSubmit hook script
• evals/evals.json — test cases for the Skill Creator eval loop
• references/setup-guide.md — explains the HITL/HOTL/HOOTL framework

To install manually if the main installer skipped it (jq missing):

brew install jq
bash ~/.claude/skills/anti-deskilling/scripts/install.sh --global

---

SKILL-template.md

A blank starting point for writing your own Claude skills. Copy it, rename the folder, and fill it in.

Structure:

• YAML frontmatter with name and description (the description drives when Claude auto-invokes the skill — write it carefully)
• Purpose section explaining what the skill improves
• Instructions with numbered steps for what Claude should do when triggered
• Constraints for positive rules and trade-offs
• Examples with good and bad trigger inputs to calibrate invocation

Place the finished file at:

• ~/.claude/skills/<name>/SKILL.md — personal, all projects
• .claude/skills/<name>/SKILL.md — project-scoped, shared via git

Use the Anthropic Skill Creator (github.com/anthropics/skills) to test and iterate with an eval loop.

---

paper-reviewer-agent.md

A ready-to-use custom subagent for journal peer review. Drop it at ~/.claude/agents/paper-reviewer.md.

Configuration:

• Model: Opus — deep reasoning needed for methodology assessment
• Tools: Read, Grep, Glob only — read-only, cannot modify files
• Permission mode: plan — no changes, exploration only

What it does: reads the full paper, summarizes the contribution, assesses methodology and claims, identifies major/minor/typographical issues, and produces a structured referee report with a recommendation. Good template to adapt for other domain-specific review tasks.

---

cpp-dev-agent.md

A ready-to-use custom subagent for C++ development. Drop it at ~/.claude/agents/cpp-dev.md.

Configuration:

• Model: Sonnet — good balance of speed and capability for implementation
• Tools: Read, Edit, Write, Bash, Grep, Glob — full access for implementation

What it does: enforces C++17/20 conventions, RAII, const-correctness, IWYU. Knows the CMake build pattern, how to run tests with ctest, and the debugging protocol (ASan → gdb → perf). Adapt the build commands and conventions section to match your actual project.

Both agent files are examples of the subagent pattern — the frontmatter controls model, tools, and permissions; the body is the system prompt. Build your own using the same structure.

---

cheatsheet.md

A single-page reference covering everything from the workshop:

• Install commands for all three tools
• Full Claude Code CLI flags table
• Operating modes and keyboard shortcuts (including /btw, Esc+Esc, Option+P)
• Configuration scopes and permission rule syntax
• Privacy env vars and data retention table (with the September 2025 opt-out note)
• Skills and subagent frontmatter reference
• Model selection guide
• Full haft /h-* command table with the FPF cycle
• Beads bd command table with solo vs. team setup
• Key links and research references

Print it or keep it open in a second window.

---

Per-Project Setup Checklist

Once the tools are installed, for each new project:

cd your-project

haft init           # decision engineering (.haft/ directory)
bd init --quiet     # issue tracking solo, OR:
bd init --team      # issue tracking with team sync

cp /path/to/starter-kit/project-settings.json .claude/settings.json
# edit to match your project's build commands

Then in Claude Code:

/plugin marketplace add steveyegge/beads
/plugin install beads

---

Privacy Reminder

Since September 2025, Anthropic trains on Free/Pro/Max data by default. If you haven't opted out yet:

claude.ai/settings/data-privacy-controls

Team and Enterprise accounts are never used for training.

---

Disclaimer

This workshop and these materials were created by Elias Karabelas in a personal capacity. NumeriCor GmbH is aware of and supportive of this event, and the NumeriCor logo is used with permission. However, this workshop is not an official NumeriCor GmbH product or service. The content, views, and recommendations expressed here are solely those of the author and do not constitute NumeriCor's advice, offerings, or business activities.

These materials are provided for educational purposes only. The author makes no warranties, express or implied, regarding the accuracy, completeness, or fitness for any particular purpose of the information contained herein. Third-party tools referenced (Claude Code, haft, Beads) are subject to their own terms and may change at any time. Use at your own risk.

---

© 2026 Elias Karabelas / NumeriCor GmbH. All rights reserved. See LICENSE for terms.
Contact: elias.karabelas@numericor.at