Every Claude Code Hack I Know (March 2026)

@kevinrose asked what IDE to use. My reply got the most engagement out of 128 answers: "No IDE. Just plan.md files and voice." Here's everything I meant by that. 1. The Moment You Have an Idea, It's /ce:plan or /ce:brainstorm The single most important thing I've learned: the moment I have an idea, it's /ce:plan. Not "let me think about this." Not "let me start coding." /ce:plan. Every time. A crazy product idea? /ce:plan. Someone posts a bug on GitHub? Copy the issue URL, paste it, /ce:plan. Error in your terminal? Screenshot it with Cmd+Shift+4, paste it directly into Claude Code with Ctrl+V, /ce:plan fix this. Claude Code accepts images - screenshots of bugs, error messages, design mockups, Slack conversations - and writes a plan from them. Here's what happens under the hood when you run it. /ce:plan launches multiple research agents in parallel. One analyzes your codebase - reads your files, finds patterns, checks your conventions. Another searches your docs/solutions/ for learnings from past bugs. If the topic warrants it, more agents research external best practices and framework docs. All simultaneously. Then it consolidates and writes a structured plan.md: what's wrong, what approach to take, which files to touch, acceptance criteria with checkboxes, patterns to follow from your own code. Not generic advice. Grounded in your codebase, your conventions, your history. /ce:work takes that plan and builds it. Breaks it into tasks, implements each one, runs tests, checks off criteria. Context gets lost? Start a new session, point it at the plan, pick up where you left off. The plan is the checkpoint that survives everything. Traditional dev is 80% coding, 20% planning. This flips it. As @jarodtaylor put it: "If you spend 80% of your time planning it with Opus and then let subagents swarm on it..." The thinking happens in the plan. The execution is mechanical. Compound Engineering is the plugin that makes this real. From @EveryInc: /plugin marketplace add EveryInc/compound-engineering-plugin I became a superfan. Then I became a contributor, the #3 contributor on GitHub, 21 commits, behind only the core team. @kevinrose introduced me to it a few weeks ago. I have 70 plan files and 263 commits on /last30days. The gap is early commits before I had this discipline. My rule now: unless it's literally a one-line change, there's always a plan.md first. 2. Get Voice-Pilled I couldn't stand voice notes before LLMs. Apple's built-in dictation made me want to throw my phone. But voice-to-LLM is different. The transcription doesn't have to be perfect because Claude Code understands context. It guesses what the mic got wrong. You can mumble, trail off, restart a sentence. Voice finally works because the listener is smart enough to fill in the gaps. Monologue (@usemonologue, from Every - same company that makes Compound Engineering) pipes speech into whatever app is focused. You talk, it types into Claude Code. WhisperFlow is great too. Pick one. I bought a gooseneck microphone for the office. I'm dictating this right now from Full Self-Driving in my Tesla, dropping off my kids. This paragraph was spoken, not typed. 3. Run Four to Six Sessions at Once This is how I actually spend my day. Four to six Ghostty windows, each running a separate Claude Code session. One is writing a plan. One is building from a different plan. One is running /last30days research. One is fixing a bug I found while testing the last thing. While /ce:plan spins up research agents in one window, I switch to another window and /ce:work a plan that's already written. While that builds, the third window gets a new bug pasted in. By the time I cycle back to the first window, the plan is done and waiting in Zed. This is why bypass permissions (next section) is non-negotiable. If every session asks "Allow?" on every action, you can't context-switch. They all need to run autonomously. Check in, react, move on. GitHub is there if you break or ruin everything. This is also why my MacBook dies in about an hour. Six Claude sessions in parallel. Just ordered the new MacBook Pro. 4. Three Settings That Change Everything Claude Code's default mode asks permission for every edit, every command. You need three config changes. "Dangerously skip permissions" (yes, that's what it's actually called). ~/.claude/settings.json: { "permissions": { "allow": [ "WebSearch", "WebFetch", "Bash", "Read", "Write", "Edit", "Glob", "Grep", "Task", "TodoWrite" ], "deny": [], "defaultMode": "bypassPermissions" }, "skipDangerousModePermissionPrompt": true } skipDangerousModePermissionPrompt: true is the key. Without it, Claude asks you to confirm every session. You can also Shift+Tab to toggle it. Credit: @danshapiro (Glowforge founder, author of Hot Seat). When I set up a friend's Claude Code, the AI actively tried to stop him from enabling this. You have to be direct. It's your computer. Sound when Claude finishes. Add to the same file: { "h