DEV Community

Cover image for I Stopped Claude Code From Breaking My Projects. Here's the Exact Setup
Creatman
Creatman

Posted on

I Stopped Claude Code From Breaking My Projects. Here's the Exact Setup

#ai #vibecoding #antigravity #claudecode

A practical guide to anti-regression AI coding with Claude Code, subagents, hooks, and Google Antigravity — from someone who was ready to quit.

Six months ago, my workflow with Claude Code looked like this: build a working prototype in an hour, spend the next three hours watching Claude systematically destroy it while "improving" things. Every developer using AI coding agents knows this pattern. You ask for a small change, and suddenly your auth module is rewritten, three tests are deleted, and there's a new dependency you never asked for.

I'm a full stack developer running my own company (CREATMAN), and I code solo most of the time. I can't afford to babysit an AI agent on every keystroke. I needed a setup where Claude Code would help me ship faster without regressing what already works.

After weeks of research, community deep-dives, and a lot of trial and error, I found a combination that actually works. This article is everything I learned — the tools, the config files, and the exact workflow I use daily.

Why Claude Code Breaks Things (It's Not What You Think)

The root cause isn't that Claude is "dumb." It's context window exhaustion.

Claude Code operates in a ~200K token window. About 80% of that gets consumed by file reads and tool results — not your conversation. By the time you're at 90% utilization, Claude literally cannot hold your project's architecture in its working memory anymore. It forgets patterns from earlier in the session, contradicts its own decisions, and starts generating code that conflicts with what it wrote 20 minutes ago.

The community calls this "context drift" and it's the #1 source of regressions. The fix isn't "write better prompts" — it's architectural.

The Stack: Antigravity + Claude Code + Subagents

Here's what I landed on after testing Cursor, VS Code, Warp, Zed, and several other setups.

Google Antigravity as the IDE

Antigravity is Google's free, agent-first IDE (a VS Code fork). I chose it for three reasons:

  1. Free. I'm already paying $100/month for Claude Max — I'm not adding Cursor on top.
  2. VS Code compatible. My extensions, keybindings, and theme transferred over in one click.
  3. Built-in browser agent. Antigravity's Gemini 3 Pro agent can autonomously open a browser, navigate your app, and test UI. No Playwright setup needed for basic visual checks.

I run Claude Code in Antigravity's terminal with my Max subscription. Gemini 3 Pro runs in the Agent Manager panel. Two AI brains, one IDE, zero extra cost.

Claude Code with Anti-Regression Config

The terminal Claude Code is where real development happens. But instead of running it raw, I set up three layers of protection.

Layer 1: CLAUDE.md — The File That Changes Everything

If you take one thing from this article, make it this. CLAUDE.md is a special file that Claude reads at the start of every session. It survives compaction. It's your project's constitution.

Here's my actual template:

# Project Name

## Architecture
- **Frontend**: React + TypeScript
- **Backend**: Python 3.11, FastAPI
- **Database**: PostgreSQL
- **Deploy**: Docker on VPS

## Key Commands
- `make dev` — Start development servers
- `make test` — Run full test suite
- `make lint` — Lint everything

## CRITICAL RULES
- NEVER delete or rewrite working tests without explicit request
- NEVER delete files without confirmation
- ALWAYS run tests after any code change
- ALWAYS do git checkpoint before large refactors
- One task at a time. Do NOT make multiple changes simultaneously
- If unsure — ASK, don't guess

## Working Style
- Plan FIRST, code SECOND. Never start coding without confirmed plan
- Small diffs. One file → tests → next file
- After every change: run tests and show results
- Use subagents for codebase research
Enter fullscreen mode Exit fullscreen mode

The CRITICAL RULES section is where the magic happens. Claude follows these instructions with remarkable consistency — because they're injected before every conversation, not buried 50 messages deep where they'd get compacted away.

Key insight from SFEIR Institute's research: 60% of Claude Code support tickets come from the "ghost context" anti-pattern — working without a CLAUDE.md. A simple CLAUDE.md resolves the issue in 90% of cases.

Layer 2: Subagents — Isolated AI Specialists

Subagents are Claude Code's most underrated feature. Each runs in its own context window and returns only a summary to your main session. This means:

  • Research doesn't pollute your implementation context
  • A reviewer can check code without knowing (or forgetting) what the planner decided
  • You can run them in parallel

I use three subagents. Create them as markdown files in .claude/agents/:

.claude/agents/planner.md — Researches the codebase and writes implementation plans. Never writes code. Saves plans to ./plans/ so other agents (and future sessions) can reference them.

.claude/agents/tester.md — Writes tests following existing patterns, runs the full test suite (not just new tests), and reports regressions.

.claude/agents/code-reviewer.md — Reviews changes for regressions, security issues, and pattern violations. Outputs severity-rated findings with file:line references.

The key is adding a reminder to your CLAUDE.md:

## Agents
- Use `planner` agent before any complex task
- Use `tester` agent after code changes
- Use `code-reviewer` agent before commits
Enter fullscreen mode Exit fullscreen mode

Without this reminder, Claude tends to do everything in the main context — which is exactly what causes context bloat and regressions.

Layer 3: Hooks — Automated Safety Nets

Hooks fire automatically on Claude Code lifecycle events. My most valuable hook blocks commits when tests fail:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash(git commit*)",
        "hooks": [
          {
            "type": "command",
            "command": "python -m pytest tests/ -x --timeout=60 || (echo {\"block\": true, \"message\": \"Tests failing.\"} 1>&2 && exit 2)",
            "timeout": 120
          }
        ]
      }
    ]
  }
}
Enter fullscreen mode Exit fullscreen mode

This creates a hard gate: Claude literally cannot commit code that breaks tests. No exceptions, no "I'll fix it later." The feedback loop forces it to fix regressions before moving forward.

The Daily Workflow

Here's how an actual development session looks:

1. Start the session

Open Antigravity. Terminal. claude. Claude reads CLAUDE.md automatically.

2. Before any complex task — plan first

> Use the planner agent to research our codebase and create
> an implementation plan for [feature]. Do NOT write any code.
Enter fullscreen mode Exit fullscreen mode

Review the plan. Question assumptions. Adjust. Only then:

> Implement Step 1 from the plan. Run tests after.
Enter fullscreen mode Exit fullscreen mode

3. Monitor context religiously

  • Check /cost periodically
  • At 60-70% context → /compact "Preserve: modified files list, test results, current plan step"
  • Switching topics → /clear

4. Use checkpoints as undo

Before risky changes: git add -A && git commit -m "checkpoint: before auth refactor"

If Claude breaks something: Esc + Esc → restore code only. Or git reset --hard HEAD.

5. Review before commit

> Use code-reviewer agent to check all changes
> Use tester agent to run the full test suite
Enter fullscreen mode Exit fullscreen mode

6. Parallel sessions for complex work

Split terminal in Antigravity:

  • Terminal 1: Claude Code — main implementation
  • Terminal 2: Claude Code — tests in parallel
  • Terminal 3: dev server

Three sessions = effectively 600K tokens of context. CLAUDE.md serves as shared memory between them.

The Antigravity Bonus: Two Brains

While Claude handles the heavy coding in the terminal, I use Antigravity's Gemini 3 Pro for:

  • UI testing via built-in browser — "Test the login flow on localhost:3000 and screenshot any errors"
  • Second opinion — paste Claude's plan into Gemini's agent panel and ask for critique
  • Documentation — Gemini is solid at generating docs from code

This dual-model approach gives you the coding power of Claude and the planning/context strengths of Gemini without paying for two subscriptions.

MCP Servers Worth Installing

For browser automation beyond what Antigravity's built-in browser offers:

claude mcp add playwright -- npx -y @executeautomation/playwright-mcp-server
Enter fullscreen mode Exit fullscreen mode

Now Claude can open browsers, take screenshots, click elements, and verify UI — all through natural language.

For GitHub integration:

claude mcp add github -- npx -y @modelcontextprotocol/server-github
Enter fullscreen mode Exit fullscreen mode

Verify everything connected with /mcp in Claude Code.

What Actually Changed

Before this setup, I'd lose 2-3 hours per day to regression whack-a-mole. Claude would "fix" one thing and break two others. I'd context-switch between debugging Claude's mistakes and actually building features.

After implementing CLAUDE.md + subagents + hooks + checkpoints:

  • Regressions dropped dramatically. The hook that blocks commits on failing tests alone is worth the entire setup time.
  • Sessions are productive longer. Subagents keep the main context clean. I can work for 2+ hours before needing to compact.
  • Recovery is instant. Checkpoints + git mean I'm never more than 10 seconds away from a working state.
  • I actually trust the output. The planner → implement → review → test pipeline catches issues before they compound.

Quick Start (15 Minutes)

If you want to try this today, grab the ready-to-use configs from the GitHub repo or set it up manually:

  1. Create CLAUDE.md in your project root with your architecture, commands, and CRITICAL RULES
  2. Create .claude/agents/tester.md — even just one subagent for testing makes a huge difference
  3. Add the commit-blocking hook to .claude/settings.json
  4. Start every complex task with "Create a plan first. Do NOT write code."

That's it. These four changes will transform your Claude Code experience more than any model upgrade or IDE switch.

Resources

I'm Nick, a full stack developer and digital architect at CREATMAN. I build backend systems, AI tools, and VPN infrastructure. The complete anti-regression setup from this article — CLAUDE.md template, all three subagents, hooks, rules, and workflow docs — is available as a ready-to-use repo: claude-code-antiregression-setup. Clone it, fill in the template, and start shipping without fear. I'm open to remote opportunities — find me on GitHub.

Top comments (0)