The 10 Claude Code Tips From the Guy Who Built It

Boris doesn't run one Claude. He runs five at the same time — five separate git checkouts of the same repo, five terminal tabs numbered 1-5, each Claude doing a different job.

Open each in its own terminal tab. Number the tabs 1-5. Run Claude Code in each. Turn on system notifications so you know when any of them needs input.

Real example: Tab 1 builds the feature. Tab 2 runs tests. Tab 3 reviews the code. Tab 4 debugs a separate bug. Tab 5 updates docs. They never conflict because each lives in its own folder.

When to skip:If you've got 8GB of RAM and a tiny project, start with 2-3 in parallel and work up.

Five wasn't enough for Boris. On top of his terminal Claudes, he runs 5-10 more sessions on claude.ai/code. He kicks them off from his phone in the morning, then pulls them into his desktop terminal when he gets to work using the --teleport flag.

The handoff is one-way: you can pull a cloud session into your terminal, but you can't push a terminal session back to the web.

Real morning workflow:Wake up. Open the Claude iOS app. Fire off 3 prompts like "audit the auth flow" or "write tests for the checkout page." Make coffee. By the time you're at your desk, three pieces of work are already done.

Boris doesn't model-shop. He uses the smartest available model with thinking on for every task. As of 2026, that's Opus 4.7 with adaptive thinking(when Boris's tweet went viral, the current model was 4.5 — the principle is the same: don't downgrade to save credits).

How to set it: In Claude Code, run /model and pick Opus 4.7. Adaptive thinking is on by default — Claude decides when to think harder based on the task.

Bigger, slower per response, but needs WAY less hand-holding. Total time-to-result is faster.

Stop doing:Switching to Sonnet to "save money" then spending 30 minutes correcting it. You burn more in corrections than Opus would've used in one shot.

Create a single CLAUDE.md file at the root of your repo. Check it into git. Every time Claude does something wrong, document the rule in CLAUDE.md so it never happens again.

Real example from Boris's team:

The compounding effect: After 3 months your CLAUDE.md is 200 lines of hard-won lessons. Every new Claude session inherits ALL of them automatically. The same mistake never happens twice.

Boris uses the Claude Code GitHub Action. During PR reviews, he tags @claude with the lesson to memorize.

Install once: Run /install-github-action in Claude Code. Approve the GitHub app.

Real PR comment example:

Claude picks it up, updates CLAUDE.md, and commits it as part of the same PR. Your team's knowledge base grows automatically while you do normal code review.

Press Shift+Tab twice. That puts Claude in Plan mode — it'll plan instead of execute. Iterate on the plan until it's actually right, then press Shift+Tab once more to flip into auto-accept mode and let Claude execute the whole thing in one shot.

The cycle: normal → plan mode (shift+tab) → auto-accept (shift+tab again) → back to normal (shift+tab).

Why it works:Most "Claude made a mistake" sessions are actually "Claude executed a bad plan." Fix the plan first. Save yourself 5 rounds of corrections.

If you do something twice in one day, it's a slash command. Drop a markdown file in .claude/commands/, check it into git, and your whole team gets it.

Example file — .claude/commands/commit-push-pr.md:

Pro move: Embed inline Bash inside the command (like git status) so Claude doesn't waste a turn fetching context. Pre-compute the info, then hand it to Claude in one go.

My take: This is the single highest-ROI tip for non-coders. You don't need to be technical to write a slash command — it's literally just plain English instructions in a .md file.

Subagents = mini specialist Claudes that handle one job really well. Boris's team has files like:

• code-simplifier.md — auto-cleans messy code after Claude writes it
• verify-app.md — has detailed end-to-end testing instructions
• build-validator.md — checks the build before merging
• code-architect.md — architecture consult
• oncall-guide.md — on-call incident response

All live in .claude/agents/. Each is a markdown file with the instructions and tool permissions for that one job.

How to invoke:Claude will pick the right subagent automatically when the task matches, or you can call one by name: "use the code-simplifier agent on this file."

Claude writes clean code 90% of the time. The other 10% breaks CI. Hooks are little scripts that fire after Claude does something — perfect for auto-formatting.

Drop this in .claude/settings.json:

Now every time Claude writes or edits a file, your formatter runs automatically. No more CI failures because Claude forgot a semicolon.

Stop using --dangerously-skip-permissions. It's lazy and unsafe. Use /permissions instead to pre-approve only the commands you actually trust.

Example allowlist (lives in .claude/settings.json):

Wildcards work. Now Claude can run any bun run testcommand without asking, but it'll still pause before doing anything risky.

Team move: Commit .claude/settings.json to git so everyone on your team gets the same allowlist.

This one wasn't in the numbered thread but Boris keeps repeating it in follow-ups: the feedback loop is everything.Don't just trust Claude's output — give it a way to check itself.

Boris uses the Claude Chrome extension. Claude opens a browser, tests UI changes live, and iterates until it's perfect. He says this single habit "2-3x's the quality" of his work.

Tell Claude: "install the Claude Chrome extension for me"