13 Proven Tricks to Double Your AI‑Assisted Coding Efficiency (From Claude Code’s Founder)

Extreme Parallelism

Boris’s core principle is “don’t let AI wait for you, and don’t let you wait for AI.” He runs five Claude instances simultaneously in separate terminal tabs labeled 1‑5 and uses system notifications to alert when a Claude needs input, effectively hiring five junior programmers while he only reviews and assigns tasks.

Marks terminal tabs 1‑5 for easy reference.

System notifications pop up when a Claude instance requires user input.

He also runs 5‑10 sessions on the claude.ai/code web interface, switching between local and cloud sessions with the & command and synchronizing context across devices via the “Teleport” feature, even checking progress on the mobile Claude app.

Model and Thinking Mode Selection

While many developers opt for cheaper, faster models like Haiku or Sonnet, Boris exclusively uses Opus 4.5 with the Thinking mode enabled. Although Opus 4.5 is more expensive and slower per request, it requires far less steering, so the total task‑completion time is lower; a single correct answer beats repeated corrections.

Team Knowledge Base – CLAUDE.md

The team maintains a CLAUDE.md file committed to a Git repository. It contains development standards such as using bun instead of npm, test commands, deployment steps, and other conventions. Claude reads this file to remember team rules and avoid low‑level mistakes, acting as a shared AI‑behaviour guide.

AI‑Driven Code Review Feedback Loop

When Claude makes a mistake in a Pull Request, the team @claude in the PR comment, prompting it to append the lesson (e.g., “don’t use enum, use string union”) to CLAUDE.md. This self‑iterating process makes the AI assistant progressively smarter and more aligned with team preferences.

High‑Level Workflow Design

Plan Mode : Most sessions start with Plan Mode (triggered by pressing Shift+Tab twice). Boris iteratively refines a plan with Claude until satisfied, then switches to “auto‑accept edits” mode, allowing Claude to execute complex tasks in a single shot.

Custom Slash Commands : Frequently used workflows are wrapped into slash commands stored under .claude/commands/. For example, /commit-push-pr automatically checks Git status, runs tests, commits code, and creates a PR, turning a multi‑step dialogue into a single command.

Sub‑Agents : Reusable agents such as code-simplifier (optimizes generated code) and verify-app (runs end‑to‑end tests) are kept as markdown files in .claude/agents/ (e.g., code-architect.md, oncall-guide.md).

Automation and Safety

PostToolUse Hook : After Claude performs a Write or Edit, a hook runs bun run format || true to automatically format code, eliminating CI formatting errors.

Permission Management : A /permissions command pre‑approves safe commands like bun test and bun run build, storing the list in .claude/settings.json. For long‑running tasks, the flag --permission-mode=dontAsk or sandbox execution lets Claude run without repeated user confirmations.

Deep Tool Integration (MCP)

Using the Model Context Protocol (MCP), Boris connects Claude to Slack, BigQuery, and Sentry. Claude can query databases, fetch error logs, and send Slack notifications, extending its role from code generator to operations assistant.

Ultimate Secret – Verification Loop

The most critical tip is to give Claude a feedback loop. When Claude sees the outcome of its code (test results, runtime errors), its code quality improves 2‑3×. Simple tasks are verified by running bash commands; complex UI interactions are handled via the Claude Chrome extension, which can click buttons and validate the UI until the experience feels correct.

Overall, Boris’s workflow emphasizes establishing clear rules ( CLAUDE.md), designing repeatable processes (slash commands, sub‑agents), providing powerful resources (Opus 4.5), and enforcing strict verification to maximize AI‑assisted development efficiency.