Abstract

AI coding assistants are now part of every developer’s toolkit - but most teams use them out of the box, with generic behavior that misses project conventions, ignores architectural decisions, and produces inconsistent output. The difference between a generic assistant and one that knows your project comes down to configuration.

This 1-hour workshop fixes that. Participants start with a bare project and progressively layer in real configuration: a shared skills framework that both Cursor and Claude Code discover automatically, MCP servers that connect assistants to external tools, and Claude Code subagents that handle complex autonomous tasks. Every addition is immediately testable – participants prompt their assistant and see the behavior change in real time.

The workshop is config-first: participants write skill definitions, wire up MCP servers, and configure assistant behavior in guided exercises, using a production-grade reference configuration (dev-config) as the blueprint.

Workshop Variants

The workshop can be tailored to the audience’s tooling:

Variant Focus Best For
Cursor-focused Skills in Cursor, MCP in Cursor, migration from legacy rules Teams standardized on Cursor IDE
Claude Code-focused CLAUDE.md, settings & hooks, skills, subagents Teams using Claude Code (CLI or VS Code extension)
Combined Shared skills & MCP config, tool-specific features, where they overlap and diverge Teams using both, or evaluating which to adopt

The syllabus below follows the combined variant. In single-tool variants, the time freed from the other tool’s section is reallocated to deeper exercises and discussion in the focused tool.

Target Audience

  • Developers and tech leads who use Cursor or Claude Code daily and want to go beyond default behavior.
  • Team leads looking to standardize AI assistant behavior across their organization.
  • Engineers evaluating how to encode project conventions, tooling preferences, and workflows into their AI setup.
  • Participants should be comfortable with a code editor and terminal; no prior experience with Cursor or Claude Code configuration is required.
  • Works well for groups of 5-15.

Syllabus

Time Phase Topic
0:00 – 0:05 – Intro & orientation – the problem with unconfigured assistants, the convergence of Cursor and Claude Code, workshop goals
0:05 – 0:20 1 The skills framework – SKILL.md anatomy, YAML frontmatter, auto-invocation vs explicit /slash commands, shared discovery across .cursor/skills/ and .claude/skills/, writing your first skill
0:20 – 0:35 2 Claude Code configuration – CLAUDE.md as persistent context, .claude/settings.json permissions model, hooks (auto-format on save, notifications), rules directory for scoped instructions
0:35 – 0:45 3 MCP servers – what MCP is and why it matters, .mcp.json walkthrough, wiring up Context7 (live docs), Linear (issue tracking), and GitHub (repo operations); shared config across both tools
0:45 – 0:55 4 Workflow skills & subagents – building multi-step workflow skills (/commit, /pr, /deps), subagents for autonomous complex tasks (code review, test generation, architecture review), when to use which
0:55 – 1:00 – Wrap-up – the dev-config setup script, sharing configs across projects via symlinks, migrating from legacy Cursor rules, open questions

Timings are approximate; each phase includes a live demo followed by a guided exercise.

What Participants Walk Away With

  • A working skills directory with at least one custom skill, discoverable by both Cursor and Claude Code
  • A configured .claude/ directory with CLAUDE.md, settings, and hooks
  • A .mcp.json wiring up at least one external tool
  • Understanding of the skills framework and how it unifies configuration across tools
  • Understanding of when to use skills vs subagents
  • Access to the dev-config reference repository and its setup script for bootstrapping future projects

Prerequisites

  • Cursor IDE (2.4+) installed, or VS Code with the Claude Code extension (or both, for the combined variant)
  • Node.js 18+ (for MCP servers via npx)
  • A GitHub account with a personal access token
  • Optional: a Linear account for the MCP exercise