Tips

The AI tricks we actually use, with the steps.

Every tip here is one we run inside our own companies. Each is a complete walkthrough — why it works, the exact steps, a snippet you can paste, and the mistakes to avoid — with the original source linked so you can check it yourself.

Beginner

Start here. Nothing to install, immediate payoff.

Pick A Project That Fits One Evening

Don't start with the big idea you came for — you'll fumble the management, blame the tool, and quit. First build three small things you personally need: each describable in one sentence, each done in an evening. Then bring out the real project.

Claude Code

Ask The Tool About The Tool

Stuck in settings, can't find a menu, button greyed out? Ask Claude Code itself — it always has access to its own current documentation, whatever version you run. For questions about the outside world, add 'check a source and give me the link'.

Claude Code

Rename Sessions Before You Drown

Sessions survive restarts — nothing burns when you close the app. But by default each is named after your first prompt, and in a week you're squinting at twenty identical grey lines. Two seconds of renaming per session buys you instant recall.

Claude Code

Start In Manual Mode On Purpose

Manual permission mode shows you a diff before every single change — accept or reject. It's slightly nagging and exactly how you learn to read what the agent does; graduate to accept-edits only once the diff-accept cycle feels like routine.

Claude Code

Steer It Mid-Run, Don't Restart

You never have to sit and watch. Type a correction while the agent works and it folds your note into the run without starting over; Esc stops it dead when you need the brakes.

Claude Code

Paste The Error, Say Nothing

When something breaks, copy the whole error message and paste it in — without commentary. Reading errors is literally the agent's job, and half of all beginner failures are people paraphrasing an error instead of showing it.

Claude Code

Point With @, Don't Describe

Type @ in the prompt field and pick the exact file — it travels to the agent with your request. Describing the file in words sends the agent searching the whole folder, which burns your plan's limits on reading things you never asked about.

Claude Code

The Four-Part First Prompt

'Make me a nice site' forces the model to invent everything, and you won't like its inventions. Name what you're making, what's in it, what you're providing, and the constraints — the more concrete the input, the more predictable the output.

Claude Code

One Project, One Folder

The folder you open is the agent's workplace and its cage — it sees nothing outside. A fresh folder per project keeps its attention on your task and your unrelated files out of reach.

Claude Code

Skip The Terminal Entirely

The Claude desktop app installs like Zoom and already contains Claude Code — Anthropic's docs literally say 'no terminal required'. The black window that scared you off vibe-coding is now optional equipment for developers, not the entry fee.

Claude Code

An Agent Does Work, Not Advice

Ask a chat how to make a landing page and you get a three-screen article. Give Claude Code the same sentence and minutes later a working page sits in your folder — it builds, checks its own result, and fixes it in a loop until the task is closed.

Claude Code

Intermediate

For people already using AI daily who want more out of it.

Make It Review Its Own Diff

'Done' is a claim, not a fact. Before accepting changes say: 'check these changes before I accept — look for real errors, not formatting'. The agent regularly finds its own bugs, and the closing phrase stops it nitpicking commas instead of logic.

Claude Code

Show A Screenshot, Skip Adjectives

Words like 'modern and minimalist' mean nothing — every ugly site's brief said that. Saw a style you like? Screenshot it into the prompt with 'in this spirit'. A picture transfers taste that three paragraphs of adjectives never will.

Claude Code

Explanatory Style For Month One

Ask for the explanatory output style and the agent narrates why it chose what it chose as it works — the black box becomes a running commentary. When the explanations start to annoy you, that's your graduation signal: switch back to default.

Claude Code

More Effort, Or A Bigger Model?

Bad result? Diagnose before you rage-quit to another tool. Skipped files, no self-check, quit halfway — that's low effort: raise the effort level. Read everything and still confidently wrong — that's the model's ceiling: escalate the model. This one question settles 95% of 'the AI got dumber' complaints.

Claude Code

Match The Model To The Chore

Limits are shared, so burning the flagship model on renaming files is how you hit the 5-hour wall by lunch. Live on the workhorse model; switch up consciously for architecture, gnarly bugs, and review — or use opus-plan mode: expensive model plans, cheap model executes, switching itself.

Claude Code

/rewind Is Your Agent's Undo

Claude Code silently checkpoints before every request — /rewind rolls files, conversation, or both back to any of the last ~100 points, surviving app restarts. It only sees the agent's own edits though: your manual changes and shell-level moves are invisible to it. It complements version control, never replaces it.

Claude Code

Two Failed Fixes Means Start Fresh

After two missed corrections the context holds three versions of what NOT to do, and every next answer gets pulled toward the wreckage. Don't correct a third time: /clear, then one good prompt carrying everything you learned. A clean session with a good prompt beats a long session of accumulated fixes.

Claude Code

One Task, One Session

The agent re-reads the whole conversation with every message you send — side questions and dead ends ride along forever, degrading answers and burning limits. /clear is free; a cluttered context is not.

Claude Code

Keep CLAUDE.md Under 200 Lines

A bloated briefing makes the agent ignore your real instructions — the docs say so directly. Test every line: 'if I delete this, will it start erring?' No? Cut it. Add a rule only after the agent makes the same mistake twice.

Claude Code

/init Writes The Project Briefing

Run /init in a new project and the agent studies your folder and writes CLAUDE.md — a briefing it re-reads at the start of every session. Without it, each session re-explores your project from zero, on your token bill.

Claude Code

Bigger Than A Button? Plan First

Plan mode makes the agent research your project and propose a written plan before touching anything — and it asks the questions you forgot to ask yourself. You edit the plan in plain words; only your accept starts the build. Redoing work costs far more than reading a plan.

Claude Code

Advanced

Wiring AI into real workflows, with the failure modes handled.

Scheduled Agents Need An Open Lid

Local scheduled tasks run only while the app is open and the machine is awake — a missed slot is skipped, not postponed. Turn on keep-awake, leave the laptop lid open, and write drift-tolerant instructions: 'if it's already past 5pm, skip today and report that you skipped'. Need lid-closed reliability? That's a cloud task — which can't see your local files.

Claude Code

Typed It Three Times? Make A Skill

The third time you type the same procedure is the signal: say 'create a skill that does this' and the agent writes the skill file itself. Then improve it in a loop — run it, note what's off, have it edit its own skill. House rules: 2nd repeated mistake → CLAUDE.md line; 3rd repeated prompt → skill.

Claude Code

Writer Session, Reviewer Session

Run two parallel sessions on one version-controlled project: one builds, the other — with fresh context — reviews. The docs are explicit about why it works: a fresh session isn't attached to code it didn't write, so it critiques honestly where the author session defends.

Claude Code

Send Noisy Research To A Subagent

Any side-investigation that would flood your session — 'which of these fifty files mentions the old pricing?' — goes to a subagent. It reads everything inside its own separate memory and returns three lines; your session's context stays clean for the actual work.

Claude Code

/context Names What Eats The Window

When the agent 'gets dumber toward evening', the window is full — /context shows exactly what by: tools, skills, messages, briefing. Then choose deliberately: /compact WITH instructions on what to preserve (it's an expensive operation — the agent re-reads everything), or the free /clear if the task is done.

Claude Code

Keys Live In .env, Never In Chat

A pasted key sits in chat history under whatever retention your settings dictate. Instead: 'create a .env file for secrets, add it to .gitignore, and verify it can never reach GitHub'. Then point the agent at the file — 'the key lives in .env, take it from there' — so the secret never crosses the chat.

Claude Code

No Connector Listed? Hand It Docs

The connector catalog is a storefront; thousands more MCP servers exist. Paste the server's documentation link and say 'wire this up, figure out the config yourself, and verify it works' — the agent edits its own config files. Only install servers from authors you trust: a malicious server is a prompt injection with a handshake.

Claude Code

Three Connectors Before All Others

Builders need exactly three to start: GitHub (the off-machine safety net), Context7 (current library docs — models otherwise code against last year's APIs), and a real search like Exa. Running a business instead? Skip the dev list and connect the tools already open in your browser tabs.

Claude Code

Let It Interview You Into A Spec

Don't write the big spec — be interviewed out of it. Have the agent grill you one question at a time, each with its own recommended answer ('your call' is a valid reply), appending every decision to spec.md. Then take that file to a FRESH session: the interview clutter stays behind, the decisions come along.

Claude Code

Comment The Diff Like A Reviewer

Open the diff panel and attach comments to exact lines, the way reviewers do in real teams — the agent reads them and fixes precisely those places. It beats describing coordinates in chat ('the second paragraph in the services block…') and it spends tokens only on what you flagged.

Claude Code

Version Control In One Sentence

Say 'take this project under version control and commit the current state as v1' — no terminal, no commands to learn. From that point every change is diffable and every broken Friday rolls back to a working Thursday. Without the baseline there is no diff counter, no rollback, no safety net.

Claude Code

Pro

Production concerns: evaluation, audit trails, running unattended.

No Verification, No Green Light

The docs' hardest rule: if you can't verify an output yourself, don't let the agent proceed until IT has. Make self-checking part of done — run it, screenshot it, list what was checked. For anything client-facing, 'it looked right' has never once been a defense.

Claude Code

Cap Research With A Deliverable

'Look into this' with no boundary is how an agent reads fifty files and stalls — the docs list boundless research among the five classic failures. Scope every investigation to a named artifact: 'a one-page summary naming the three files to change'. The deliverable is the fence.

Claude Code

Tell Compaction What To Keep

When a long conversation gets squeezed, the agent decides what survives — and may drop exactly what you needed. One line in CLAUDE.md fixes it forever: 'when compacting, always preserve the list of changed files and my requirements'. Works in every session, no reminders, and /compact accepts the same guidance inline.

Claude Code

Compact Early, On Your Terms

Auto-compaction fires when the window is nearly full — which is after quality already sagged, and the automatic summary guesses at what mattered. Compact earlier and by hand, or ask the agent to lower its auto-compact threshold, so the squeeze happens while the session is still sharp.

Claude Code

Artifacts Demo It, Hosting Ships It

'Publish this as a shareable page' gets you a live URL in a minute — it republishes to the same address as you iterate, which is perfect for showing a client work in progress. Know the box you're in: one page, no backend, 16MB, private or public. Forms, logins, and payments are your cue to move to real hosting.

Claude Code

Run Your Agent From Your Phone

Dispatch turns your pocket into a task queue: text the agent like you'd text an employee, and it opens a session on your machine and works. The prerequisites all live at your desk: app open, keep-awake on, mobile notifications enabled, browser actions pre-approved, and a permission mode chosen for sessions it spawns — decide all of it before you leave.

Claude Code

Set The Training Toggle On Purpose

On personal plans one privacy setting decides both whether your sessions train future models and how long they're retained — years with it on, weeks with it off. Anthropic documents this openly; most users have simply never opened the page. Whatever you choose, choose it — and keep client PII and NDA material out of working folders regardless.

Claude

Bypass Mode Belongs In A Sandbox

No-questions mode is handing a first-day junior the office keys, the company stamp, and the bank account — the docs restrict it to isolated environments for a reason. On a rented server it's how a database disappears 'with no undo': GitHub restores your code, never your data.

Claude Code

MCP Replies Bill Your Context

Connecting servers is nearly free now — tool descriptions load on demand. The cost moved to responses: whatever a server returns lands in your window whole (warning at 10k tokens, truncation at 25k). A CLI tool lets the agent filter output before it touches context; an MCP server doesn't. Let the agent pick per job — it knows both catalogs.

Claude Code

Two Locks Keep A Sandbox Honest

Sandboxed code execution is safe because of two locks: no filesystem outside the box, no network egress. Open the network and a prompt-injected agent can exfiltrate whatever it just read; keep egress shut and even a successful attack has nowhere to send the loot. Use the domain allowlist for the few hosts a job genuinely needs.

Claude Code

Grant Rights, Don't Write Pleas

'Never touch production' in a rules file is a wish the agent can overlook. A scoped key that physically cannot reach production is a setting it cannot ignore. Every boundary you actually care about belongs in credentials and permissions, not in polite prose.

Claude Code