Your AI coding agent scores 10/100. Here's what it's missing.

Nerviq — Mon, 06 Apr 2026 13:23:05 +0000

After testing 2,431 checks across 8 AI coding platforms on real repos, I found a clear pattern: most projects use barely 10% of what's available.

I built Nerviq — a zero-dependency CLI that audits your AI coding agent setup and scores it 0-100.

npx @nerviq/cli audit

Most projects score 10-20 out of 100. After running setup, they jump to 60-80+.

The Top 10 Things You're Probably Missing

1. Instructions file (Critical)

Every AI coding platform has one: CLAUDE.md, AGENTS.md, .cursorrules, GEMINI.md. Without it, the agent doesn't know your build commands, code style, or project rules.

2. Architecture diagrams (73% token savings)

A Mermaid diagram gives your agent the project structure in a fraction of the tokens that prose requires.

3. Hooks > instructions (100% vs 80% compliance)

Written instructions are advisory (~80% compliance). Hooks are deterministic (100%). Auto-lint after every edit. Every time.

4. Verification commands

This is the single highest-leverage thing you can do. — Anthropic Best Practices

Agents perform dramatically better when they can verify their own work.

5. Deny rules

Your AI agent loads .env files automatically. Without deny rules, it can read secrets. This is the most common critical finding.

6. Custom commands

Stop typing the same prompts. Create /test, /deploy, /review as reusable commands.

7. Cross-platform drift

If you use Claude + Cursor + Copilot, their configs probably conflict. nerviq harmony-audit catches this automatically.

8. Permission profiles

Most setups either allow everything or block everything. There are better profiles: read-only, suggest-only, safe-write, power-user.

9. MCP configuration

MCP servers extend what your agent can do, but >10 servers or >80 tools degrades performance.

10. Stack-specific checks

Python, Go, Rust, Java, Ruby, PHP, .NET, Flutter, Swift, Kotlin — each has specific agent configuration needs.

New: Auto-fix

nerviq fix                    # Show what's fixable
nerviq fix claudeMd           # Fix a specific check
nerviq fix --all-critical     # Fix all critical issues

Every fix shows the score impact: Score: 4 → 16 (+12).

Safety first

nerviq audit — reads only, writes nothing (default)
nerviq suggest-only — markdown report, no file writes
--dry-run — preview changes
--config-only — only touches config files, never source code
Every write creates rollback artifacts

Try it

npx @nerviq/cli audit              # Quick scan: score + top 3
npx @nerviq/cli audit --full       # All checks + confidence levels
npx @nerviq/cli suggest-only       # Share results with your team
npx @nerviq/cli harmony-audit      # Cross-platform alignment

Zero dependencies. No API keys. Runs locally. AGPL-3.0.

GitHub: github.com/nerviq/nerviq
npm: npmjs.com/package/@nerviq/cli
Website: nerviq.net

972 Claude Code Techniques: What I Learned Building the Most Comprehensive Catalog

Nerviq — Mon, 30 Mar 2026 19:38:22 +0000

After 39 research cycles, I cataloged 972 Claude Code techniques and tested 773 of them with real evidence. Here are the top 10 things most developers are missing.

The Problem

Claude Code is incredibly powerful - 74 releases in 52 days, hundreds of features. But most projects barely scratch the surface.

I built a CLI that scores your project:

npx claudex-setup

Most projects score 10-20 out of 100. After running setup, they jump to 70+.

Top 10 Things You're Probably Missing

1. CLAUDE.md (Impact: Critical)

Claude reads this file at the start of every session. Without it, Claude doesn't know your build commands, code style, or project rules.

Run /init to auto-generate one from your codebase.

2. Mermaid Architecture Diagrams (Saves 73% Tokens)

A Mermaid diagram in CLAUDE.md gives Claude your project structure in a fraction of the tokens that prose would require.

3. Hooks > CLAUDE.md Rules (100% vs 80%)

CLAUDE.md instructions are advisory (~80% compliance). Hooks are deterministic (100%). Auto-lint after every edit. Every time. No exceptions.

4. Custom Commands

Stop typing the same prompts. Create /test, /deploy, /review in .claude/commands/.

5. Verification Loops (The #1 Best Practice)

Claude performs dramatically better when it can verify its own work. Include test commands in CLAUDE.md.

This is the single highest-leverage thing you can do. - Anthropic Best Practices

6. Path-Specific Rules

Different conventions for frontend vs backend files in .claude/rules/.

7. XML Tags for Structured Prompts

Use <constraints>, <validation> in CLAUDE.md for unambiguous instructions. 30% quality improvement (Anthropic-measured).

8. Custom Agents

Specialized subagents for focused tasks: security-reviewer, test-writer in .claude/agents/.

9. Skills (On-Demand Knowledge)

Unlike CLAUDE.md (loaded every session), skills load only when relevant. Run /fix-issue 1234 and Claude handles the rest.

10. MCP Servers

Connect Claude to your database, ticket system, or Slack via MCP configuration.

Try It Now

npx claudex-setup        # Audit (10 seconds)
npx claudex-setup setup  # Auto-fix

Free, open source, zero dependencies.

GitHub: github.com/DnaFin/claudex

Built from CLAUDEX - 972 verified Claude Code techniques, 773 tested with real evidence.

DEV Community: Nerviq