Part 1 of 5
Set Up Right (Solutions 1-4)
Switch Model and Effort Depending on the Task
Stop using one model for everything. Big models are for hard thinking: planning, writing, debugging. Small models are for grunt work: renaming files, quick lookups, simple edits.
In Claude Code, type /model to switch. Match the horsepower to the job and you save money AND get better results, because big models get bored and overcomplicate simple tasks.
Use Desktop Apps Unless You're an Advanced Coder
The terminal looks cool. It's also where beginners get stuck. The Claude desktop app and Claude Code desktop app give you the same power with buttons, file previews, and readable output.
Use the terminal when you actually need it, not to feel like a hacker.
Save Your API Keys in an ENV File
Never paste an API key into a chat, a script, or a prompt. Create one file called .env, put every key in there, and tell AI "the key is in my .env file."
AI reads it without ever exposing it. One file, all your keys, nothing leaked in a screenshot or a shared chat.
Install Destructive Command Guard
AI can delete files. It can wipe a database. Not because it's evil, because it's fast and confident. One rm -rf or DROP TABLE and hours of work is gone in seconds.
Destructive Command Guard (dcg) is a free open-source tool that fixes this. It sits between the AI and your computer as a safety hook. Every command the AI wants to run gets inspected BEFORE it executes. If it matches a dangerous pattern (deleting folders, wiping git history, dropping database tables, formatting disks), it gets blocked and you get asked first. It even catches destructive code hidden inside scripts, not just direct commands.
It plugs straight into Claude Code, Codex, Cursor, Copilot, and Gemini CLI. And it's fast: checks happen in under a millisecond, so you never feel it.
How to install:Go to the repo below and tell Claude: "install this for me." Or run the one-line installer from the README.
The installer auto-detects your setup and configures the hooks for you. Five minutes of setup saves you from the one mistake you can't undo.
Part 2 of 5
Keep the Context Clean (Solutions 5-9)
Ask for a Handoff File at 50% of the Context Window
AI gets dumber as the chat gets longer. Don't ride a conversation until it dies. At the halfway point, say: "create a handoff file with everything the next chat needs to continue this work."
Start fresh, paste it in, and the new chat picks up exactly where you left off, at full intelligence.
Start From Scratch After Two Bad Outputs in a Row
Two bad answers means the chat is polluted. The AI is now anchored to its own mistakes, and every correction digs the hole deeper.
Don't argue with it. Open a new chat, rewrite the request with what you learned, and you'll usually nail it on the first try.
Keep Your CLAUDE.md Under 100 Lines
CLAUDE.md is the instruction file Claude reads at the start of every session. When it's 400 lines, Claude skims it and ignores half of it.
Keep only the rules that matter on every single task. If a rule only matters sometimes, move it to a separate file and point to it.
Keep Your memory.md Under 200 Lines
Same problem, different file. Memory is useful until it's a junk drawer. Keep the index short, push details into topic files, and delete anything stale.
A lean memory gets read. A bloated one gets skimmed.
Use Subagents for Big Jobs So One Chat Stays Clean
Big job? Don't do it all in one chat. Tell Claude to "use subagents" and it spins up workers that each handle a piece (research this, build that, test the other thing) and report back.
Your main chat stays clean and high-level instead of drowning in details. That's how you do big work without hitting the context wall.
Part 3 of 5
Prompt Like a Pro (Solutions 10-15)
Screenshot the Error, Don't Describe It
"It's broken" tells AI nothing. Your description of an error is a bad translation of it. Screenshot the actual error and paste the image in.
AI reads the exact message, the line number, the stack trace. Ten seconds, and it fixes in one shot what would've taken five rounds of guessing.
Voice Dictate Long Prompts Instead of Typing
Long prompts get better outputs, but nobody types long prompts. So talk. Use your keyboard's dictation key or any voice tool and ramble for two minutes about what you want.
Messy speech beats three tidy sentences, because AI is great at organizing your mess and terrible at reading your mind.
Treat the First Output as a Draft
The first answer is never the best answer. It's the average answer. Pros treat output one as a starting point: "make it punchier," "cut the fluff," "now make it sound like me."
Two or three rounds of pushing is where the quality lives. People who accept the first output are the ones complaining AI is mid.
Ask for Three Options, Pick One, Then Go Deep
Don't ask for "a design" or "a headline." Ask for three different directions. Pick the winner, then say "go deep on this one."
You skip the round of "no, not like that" entirely, because you chose the direction instead of hoping AI guessed it.
Ask "What Could Make This Go Wrong?"
Before you ship anything AI built (a plan, code, an email), ask it: "what could make this go wrong?" It will find holes in its own work that it would never volunteer.
It's a free red-team review, and it catches the stuff you'd have found out the hard way.
Make AI Verify Its Work Before Saying It's Done
"Done" from an AI means "I wrote something," not "it works." Change the rule: "run it and show me the output before you say done," "screenshot the page," "test it first."
When verification is part of the job, you stop shipping broken stuff you didn't check.
Part 4 of 5
The Free Media Toolkit (Solutions 16-18)
Use yt-dlp to Download Videos From the Internet
Free command-line tool, downloads video from almost anywhere. Install once, or just tell Claude "install yt-dlp and download this video" with a URL.
That's your competitor research and your repurposing pipeline, no sketchy download sites.
Use ffmpeg to Extract the Audio
Free tool, converts anything to anything. Install it once, then tell Claude "extract the audio from this video."
One command turns any video into an mp3, ready for the next step.
Use Whisper to Transcribe Any Audio
OpenAI's free transcription model, runs on your own computer. Tell Claude "install Whisper and transcribe this audio."
Chain all three tools and any video on the internet becomes a text file: download it (yt-dlp), strip the audio (ffmpeg), transcribe it (Whisper). That transcript becomes scripts, posts, guides, whatever you need.
Part 5 of 5
Make It Sound Like You (Solutions 19-20)
Create a Brand Voice Document
AI defaults to sounding like everyone. Fix it once: a single doc with how you talk, words you use, words you'd never use, and 3-5 examples of your real writing.
Then every prompt starts with "follow my brand voice doc." Design and writing outputs both jump instantly, because AI finally knows who it's writing as.
Use the Impeccable Skill to Build Websites
Default AI websites all have the same generic look. Skills are add-on instruction packs for Claude Code, and impeccable is one built specifically for frontend design: real hierarchy, real polish, no template smell.
Install it, then just say "use the impeccable skill" when you build. Same prompt, dramatically better site.
All Resources
Prefer watching over reading?
Watch the Reel on Instagram →