8 Claude Code Hacks That Supercharged a Lyft Co‑founder’s AI Coding Productivity

Introduction

Matt Van Horn, co‑founder of Lyft and founder of the smart‑oven startup June, published a detailed March 2026 post titled “Every Claude Code Hack I Know,” which amassed 620 k views, 7.3 k saves, and 1.8 k likes. He also open‑sourced the /last30days Claude Code skill that aggregates recent AI‑relevant content.

Hack 1 – Plan First

Van Horn’s first principle is to never let Claude write code directly. He invokes the /ce:plan command (or prompts “plan first, don’t code”) so Claude reads files, analyses architecture, and lists steps. After the plan is reviewed, Claude executes the code. Benefits listed in a highlighted box include reducing drift (Claude won’t take unintended actions), cutting costs (a clear plan avoids 80 % of wasted effort), and enabling auditability (the plan can be edited before execution). An engineering lead is quoted: “If you have a precise, detailed plan, Claude never goes off‑track.”

Hack 2 – Voice Input

Van Horn recommends voice‑input tools such as Monologue because speaking is three times faster than typing and naturally yields richer descriptions. He contrasts a four‑word typed prompt “fix the login bug” with a spoken, detailed description of the bug, noting that voice input provides ten‑fold more context, which leads to more accurate Claude outputs.

Hack 3 – Parallel Sessions with Ghostty

Instead of treating Claude as a single‑threaded assistant, Van Horn runs 4‑6 concurrent Claude Code sessions in the Ghostty terminal. Each window handles a distinct task (frontend component, backend API, tests, code review). He colors each session with the /color command for instant visual identification. Resource comparison shows VS Code running two Claude projects consumes 8.2 GB RAM and 12 % CPU, whereas Ghostty uses 487 MB RAM and 1 % CPU for the same load; eight Ghostty sessions use only 1.9 GB RAM with negligible impact.

Hack 4 – Configure CLAUDE.md

Van Horn stresses the importance of a CLAUDE.md file as a system‑level prompt loaded on every start. The file contains project context (tech stack, architecture decisions), behavior directives (e.g., “plan before execute,” “don’t modify unread code”), code style guidelines (testing requirements, Git commit format, PR template), and security boundaries (files or commands that must not be touched). Combined with the dangerously‑skip‑permissions flag in container environments, Claude can run autonomously, requiring only result review.

Hack 5 – Real‑Time Research with /last30days

The /last30days skill scans eight platforms (Reddit, X, YouTube, TikTok, Hacker News, Polymarket, etc.) for the past 30 days, ranks content by social signals, synthesizes the top material, and feeds it to Claude. This mitigates the lag in large‑model training data, ensuring Claude has up‑to‑date knowledge of new APIs, versions, and best practices.

Hack 6 – Meeting‑Driven Development

Using the AI meeting‑recording tool Granola, Van Horn records product discussions, extracts a structured transcript, and pastes it into Claude Code. Claude then generates a development plan based on the conversation. This preserves full context and decision rationale, eliminating the need for separate requirement documents.

Hack 7 – Markdown Planning for Non‑Code Projects

Claude Code can also handle structured‑thinking tasks. Van Horn demonstrates planning a family trip by creating a Markdown file that lists constraints (budget, time, member needs). Claude reads the file in Plan Mode and outputs a detailed itinerary, complete with reservation times, routes, and restaurant recommendations. Unlike ChatGPT, Claude can read/write local files, iterate, and version the plan with Git.

Hack 8 – Remote Development with Mac Mini + Telegram

The final hack sets up a Mac Mini as a 24/7 Claude Code server. Claude Code’s new Channels feature (released March 20) allows messages from Telegram or Discord to trigger Claude actions on the Mac Mini. Sending a Telegram message starts Claude work remotely, freeing MacBook resources, supporting long‑running background tasks, and providing always‑available AI assistance.

Editorial Insight

Van Horn argues that the real value lies in the methodology, not the individual tools. He summarizes the workflow as:

Input: Voice input + meeting recordings for richer context.

Planning: Plan Mode + CLAUDE.md to prevent drift.

Execution: Parallel sessions + Ghostty for maximum throughput.

Knowledge: /last30days for real‑time research.

Remote: Mac Mini + Telegram for ubiquitous access.

The article’s popularity (7300+ saves) underscores a community need for systematic AI‑coding practices rather than just better models.