Install Testing Dependencies

Run npm install with Jest, ts-jest, testing libraries, and supertest. These provide the foundation for reliable unit and integration tests.

Configure Jest with 80% Coverage Thresholds

Create jest.config.ts with coverageThreshold set to 80% for branches, functions, lines, and statements. Use collectCoverageFrom to target source files and exclude tests and bootstrap code.

Add Test Scripts to package.json

Create test and test:coverage scripts that run Jest locally and in CI. This ensures coverage checks are enforced automatically.

Write Failing Tests First

Before asking AI for implementation, write a test that describes the behavior you want. Be specific about inputs, outputs, and edge cases.

Ask AI to Make Tests Pass

Show the failing test to the AI and ask it to write code that makes the test pass. The test becomes the specification the AI must satisfy.

Run Coverage Check and Merge

Execute test:coverage before merging. If coverage drops below 80%, the build fails. No override, no exceptions — the gate is automatic.

AI generates plausible code fast. Tests are the only signal that it's actually correct. Here's how to set up Jest, coverage thresholds, and CI gates before you touch a single line of AI-generated code.

Don't Write a Line of AI Code Until Your Tests Are Green

I just finished The Pragmatic Programmer. Published in 1999 — no AI, no cloud. It hits harder now than when it was written. Here's why every principle in that book matters more, not less, in the age of AI.

The Pragmatic Programmer Is More Relevant Now Than When It Was Written

Start a New Chat for Each Task

Create a fresh chat for every distinct task instead of piling requests into one long conversation. This keeps Claude's attention focused and prevents context degradation.

Use Agents for Sub-tasks

Ask Claude to break tasks into sub-agents and handle each separately. This keeps your main context lean and prevents the main thread from being contaminated by rabbit holes.

Compress at Natural Breakpoints

When a task completes or context reaches 50% full, ask Claude to summarize decisions and state, then start a new chat with just that summary. This preserves signal and ditches noise.

Use Plan Mode Before Execution

Request a plan before implementation. Review it, flag issues, then approve. This catches wrong direction early instead of undoing hundreds of lines of code.

Apply Token-Efficient Rules Globally

Copy or fork Claude Token Efficient Rules into your configuration to reduce verbose output and get tighter, more direct responses by default.

Use Serena for IDE-level Workflows

Register Serena as an MCP server in your IDE to manage context handoffs cleanly without pasting large code chunks into Claude's chat.

Leverage VoltAgent Subagent Library

Use pre-built subagents from VoltAgent Awesome Claude Code Subagents instead of building agents from scratch. They are already scoped correctly.

You're not hitting Claude's limits because of the plan — you're hitting them because of your workflow. Here are the 7 techniques I used to dramatically increase my output on the $20 plan.

Ai-tools

Don't Write a Line of AI Code Until Your Tests Are Green

The Pragmatic Programmer Is More Relevant Now Than When It Was Written

I'm on Claude's $20 Plan. Here's How I Made It Feel Like I Have No Limits.