The Claude Code tip economy has a supply problem. There are 27-tip posts, 30-tip posts, and 50-tip posts, most of them summarizing each other, and the count only goes up.
Now, something that needs to be said: a growing share of the tips are either already obsolete, because the tool absorbed them as defaults, or were never written for someone with your workload in the first place. A smaller share is worse than useless, because following it makes Claude Code perform worse. You can safely ignore most manual context micromanagement, most MCP maximalism, and almost all of the parallel-session choreography, and lose nothing.
This post is only about what to drop. The habits I kept are in my Claude Code workflow tips.
Why do most Claude Code tips miss?
A tip is a snapshot of one person's workflow at one moment in a tool that ships updates weekly. Claude Code is a terminal-based coding agent that reads your codebase, plans changes, edits files, and runs commands, with a CLAUDE.md file as persistent project memory. The product changes fast enough that advice written in February describes a different tool by July.
The second reason is volume. A YouTuber mass-producing demo apps has a real use for running four sessions across git worktrees. A developer maintaining client projects has a real use for maybe one of those sessions, supervised. The tip is true for the author and expensive for you.
I've watched this pattern before, from the other side. At a client, I owned architecture for 30+ concurrent client projects, and the recurring failure mode was teams adopting whatever workflow a conference talk demoed that month, regardless of whether their delivery volume justified it. The fix there was boring process discipline, and it's the same fix here.
Which tips has Claude Code already absorbed?
The biggest category of ignorable advice is manual context babysitting. Claude Code auto-compacts the conversation when the window fills, so the "watch your percentage and compact at the right moment" ritual from early 2026 tip lists is now mostly the tool's job. Running /clear between unrelated tasks is still worth doing, because starting clean beats summarizing a mess, but the anxiety-driven monitoring can go.
The thinking ladder went the same way. For most of 2025, every tip list carried the same table: "think" < "think hard" < "think harder" < "ultrathink", each phrase buying a bigger reasoning budget. The current model configuration docs say Claude Code passes "think", "think hard" and "think more" through as ordinary prompt text. Only ultrathink is still recognized, and only for the turn you type it in. The actual control is /effort, with levels from low to max, and it holds for the whole session.
Typing "think harder" into a prompt now does about as much as typing it into an email to a colleague.
The reusable-prompt tips aged too. Advice about keeping a folder of prompt templates to paste into sessions predates skills, which are files the model loads on demand when your task matches their description. If you're still pasting the same instructions twice, that prompt should already be a command. And the tips that tell you to build a .claude/commands/ folder are a step behind: the skills docs say custom commands have been merged into skills. Old command files keep working, but a skill can also carry supporting files and get picked up automatically when it's relevant.
MCP maximalism aged badly as well. Early advice said connect everything: database, browser, project tracker, design tool. Then every connected server dumped its tool definitions into your context and quality dropped. Claude Code now loads tool definitions dynamically when needed, and Anthropic's own best practices call CLI tools like gh the most context-efficient way to reach external services. The current best practice is closer to a short allowlist than a collection.
Which tips make Claude Code worse?
Some advice was never harmless, and it still circulates.
The first one sounds responsible: put every rule in CLAUDE.md, since it loads in every session. Anthropic's best practices page says the opposite, and in plain words: bloated CLAUDE.md files cause Claude to ignore your actual instructions. Their test for each line is whether removing it would cause Claude to make mistakes. Build commands Claude can't guess pass that test. "Write clean code" doesn't. A procedure that only matters for some tasks belongs in a skill, which only loads when it's needed.
Its companion tip is writing IMPORTANT and YOU MUST in capitals on every rule. The same page recommends emphasis on the one instruction Claude keeps skipping, because when many lines are emphasized, none of them stands out anymore.
Then there's the permissions flag. Plenty of 2025 tip lists told you to start every session with --dangerously-skip-permissions so the approval prompts would stop. The permission modes docs list that mode as meant for isolated containers and VMs only. Since v2.1.283, interactive sessions start in auto mode instead, where a separate classifier model reviews actions and blocks the risky ones, like scope escalation or commands driven by hostile content. The annoyance the flag solved is mostly handled now, and the risk it created is still there. I wrote about what a prompt injection can do with an unsupervised agent in the sandbox post.
Plan mode for everything is the gentler version of the same mistake. Planning pays off when you're unsure of the approach or the change spreads across several files. The docs' own rule: if you could describe the diff in one sentence, skip the plan. Forcing a plan on a typo fix adds a review step to a change you already understood.
Which tips were never for you?
Some advice is accurate and still worth ignoring, because it solves a scale problem you don't have.
Parallel sessions with git worktrees, fleets of pre-configured subagents, notification hooks so you can run three tasks while away from the desk, spending caps for agents running inside CI pipelines. All of it is real and documented, and all of it earns its keep at volume. All of it also adds configuration and supervision overhead that a solo developer or a small agency team will spend more time maintaining than saving.
My rule from mentoring a team of five developers applies unchanged, and it is the kind of call fractional CTO work exists to make: nobody gets a process upgrade until the current process is the bottleneck. If your sessions still drift because your task framing is vague, a second parallel session gives you two drifting sessions. I keep a single supervised session for WP-AutoInsight, a plugin with 6,000+ downloads, and the constraint is review capacity, every single time. Session count has room to spare.
That constraint doesn't show up in tip lists, because it doesn't have a slash command.
What if you never open a terminal at all?
An increasing share of the people asking me about Claude Code tips shouldn't be using Claude Code. Anthropic now ships Claude Cowork, an agentic desktop app aimed at knowledge work rather than programming, and Claude Desktop covers a lot of what non-developers were forcing through the CLI a year ago. For a marketer running content pipelines or an operations person automating reports, terminal tips are trivia.
What transfers across every surface is the working method. Plan before executing. Describe the problem and the constraints instead of dictating steps. Review output before it compounds into more output. Those habits matter identically in a terminal session, a Cowork task, or an n8n workflow versus a full agent setup, and choosing the right surface for the job matters more than any trick on the wrong one.
What's left after you drop all this?
Less than the tip lists suggest. I keep that list in a separate post, Claude Code workflow tips from 18 months of daily use, because what to do and what to stop doing are different questions.
One rule belongs here, though, because it's the filter for everything above: check any tip against the official docs at code.claude.com before adopting it. Tip lists routinely mix real commands with misremembered feature names, and each round of AI-summarized summaries adds confidence while removing accuracy. The thinking ladder is the proof. Anthropic documented it once, changed it, and the tip lists kept repeating the old table for months.
If it's in no official doc, it doesn't exist.
People getting real leverage from Claude Code read the same lists as the people collecting tricks. The leverage comes from knowing which ten percent applies to your actual workload, and having the discipline to ignore the rest.
Deciding what your team should automate, which surface it belongs on and which shiny options to skip is the judgment call I handle in fractional CTO and AI advisory engagements. The tools change monthly. The decision framework doesn't. Start here.