If you've been using Cursor or Claude Code for any length of time, you know how this goes: paste a system prompt, spend the first week coaching the AI through your preferred patterns, get inconsistent output anyway because nothing is actually encoded anywhere permanent. Repeat on every new project. It's not a tooling problem—it's a structure problem.

The Core Insight: Format Is Everything

The author behind this pack—who publishes under Founder Prompts on DEV.to—spent weeks encoding their most-repeated patterns into files built to the exact format these tools expect. We're talking YAML frontmatter in .md agents, glob patterns for rule matching, shell script hooks with specific exit codes, and skill directories with SKILL.md files. These aren't prompts you paste at conversation start; they modify how the toolchain itself behaves. Drop them in, and they're installed—no configuration step required.

Four Subagents That Do Real Work

The pack includes four specialized subagents stored as .md files in .claude/agents/. The code-reviewer agent structures its output into correctness bugs, OWASP-tagged security issues (with a constraint forcing every finding to include a Top 10 reference or explicitly note "not mapped"), data-loss risks, and performance problems—plus a merge verdict so you're not left hanging. The test-generator detects your framework (Vitest, Jest, pytest) and existing conventions, then produces zero-stub tests with real assertions—no placeholder // TODO blocks allowed. The dependency-auditor runs native audit tools when available but falls back to manual analysis with named CVE IDs rather than vague "potential vulnerability" notes. And the migration-runner is the heavy lifter: it rewrites ADD COLUMN NOT NULL operations into a three-step sequence (add nullable, backfill in batches, add constraint separately) to avoid table locks on production databases with millions of rows.

Hooks That Actually Block Bad Edits

Claude Code hooks run shell scripts at specific points in the edit lifecycle. A PreToolUse hook fires before any Edit, Write, or MultiEdit call—and if it exits with code 2, Claude Code blocks the edit. The pack includes a pre-commit-lint hook that runs ESLint on JS/TS files, Ruff on Python, or cargo clippy on Rust, then exits 2 if problems exist. The non-obvious detail: it handles uninstalled linters gracefully by exiting 0 silently rather than breaking your workflow with false failures. There's also a post-edit-test hook that mirrors this behavior after edits but never blocks—only informs.

Skills as Protocols, Not Instructions

Skills are newer Claude Code primitives: folders in .claude/skills/ containing a SKILL.md file. The pack includes two. The long-running-tasks skill implements checkpoint-based execution—you write _state.json before starting multi-step pipelines, update it after each step, and resume from the last completed index on crash or restart. This is how you run hour-long operations through Claude Code without losing progress when something goes wrong. The api-integration skill encodes a production-grade protocol covering credential lookup from environment variables (never hardcoded), connection timeouts with exponential backoff and jitter, Retry-After header handling, idempotency keys on POST requests, circuit breaker state, and webhook signature verification.

What You Get for the Download

The full pack is 12 files: 4 subagents, 4 Cursor rules, 2 hook scripts plus their JSON merge snippets, and 2 complete skill folders with templates. Each file represents 30 to 60 minutes of real iteration—writing a first version, testing against actual inputs, finding failure modes, then encoding the fixes as explicit constraints in the system prompts.

Key Takeaways

  • Subagents delegate based on description fields; write them like you write API documentation
  • PreToolUse hooks that exit code 2 block edits; exits 0 silently when nothing is configured
  • Skills encode protocols, not instructions—checkpoint state files survive crashes and restarts
  • The migration-runner's three-step ADD COLUMN pattern prevents production table locks on large datasets

The Bottom Line

If you're using Claude Code or Cursor daily and still coaching it from scratch on every project, you're leaving time on the table. This pack won't make you a better engineer—but it'll make your AI assistant behave like one you've already trained. That's worth the download alone.