Will Claude Code Respect Your Boilerplate? 6 Tests

Brilliant agent, feral repo

Three months of daily Claude Code ended with an uncomfortable pattern: UI correct, agent says done, code compiles — and POST /api/order returns 500 on a coupon feature. "AI coding agents are interns with perfect syntax and zero accountability. They can write, but they cannot verify reality." The 13-agent workflow that broke sharpens it: ignored instructions, autonomous tool runs past guardrails, shallow reasoning, fabricated file reads — on a $200/mo Max plan running hooks and phase gates. Same root cause in both: the agent was fine, the repo gave it nothing to hold onto.

The scorecard: rate any boilerplate in 60 seconds

Score your current starter out of six. Most score two.

BoringPush was built to score six: a written agent rules file plus reusable skills (including a mobile-session guard), tri-database parity across SQLite/MongoDB/Supabase behind identical behavior, thin API routes with the access check first, and the gate every turn ends at. Ship from anywhere on this Next.js stack and a phone session gets the same fences as a desktop one — the rules live in files Claude reads, not in habits humans forget.

The Monday test: audition any repo in an afternoon

1. Read test. Ask the agent "where do orders get created?" It should cite the real route and data files with line numbers, not narrate from memory. Vague answers = no locality to exploit.
2. Field test. "Add twitterHandle to users." Watch: one place or five? Same-shape providers turn it into schema + migration + done. Scattered models turn it into an afternoon.
3. Slop test. "Add rate limiting." If the diff touches fifteen files and installs a package, the repo failed. If it touches one guard file and zero packages, it passed.
4. Refund test. git stash the session. Clean tree in one command? Good — small blast radius. Hour-long untangle? The repo has no seams.

FAQ

Claude Code vs Cursor vs Copilot — does the repo choice change?

No. The scorecard is tool-agnostic: instruction files, thin routes, and gates help every agent.

Pick the editor you like; fix the repo either way.

What goes in AGENTS.md vs regular docs?

Rules the agent must obey (never install, additive migrations only, thin routes) go in AGENTS.md. Explanations go in docs.

Agents follow constraints; they skim prose.

My repo scores 2/6. Rewrite or retrofit?

Retrofit in this order: lint gate first (one afternoon, catches everything), thin the hottest three routes, then add the instruction file. Migrations and parity come last — they're structural.

Do hooks and subagents replace repo conventions?

No — the broken 13-agent workflow ran 89 hooks and still fabricated output. Hooks enforce process; conventions remove the need for heroics.

You want both, but conventions first.

Why three databases instead of one good one?

Because "which database" is the most rewritten decision in SaaS. Parity makes it a one-variable deployment detail, so neither you nor the agent ever rewrites queries to switch.