Claude Code Advanced Guide: 10 Tips to Stop Approving Manually

Why Is Your Claude Code Still in "Manual Mode"?

At 9:40 PM on a Tuesday, a construction crew lead's phone buzzed with the 40th approval request. The scheduling app he'd built with Claude Code was already usable, but now he wanted to add three features: a materials list for each project, a timesheet, and automatic next-week scheduling every Friday. Since finishing dinner, he'd been tapping "Approve" every few seconds until the screen showed "Usage limit reached, resets after midnight"—only one item on his list was done.

According to analysis from the YouTube channel The Benmore Brief, the problem isn't Claude Code itself, but that the user is still stuck in "beginner mode": sitting at the computer, staring at every step. The channel's 10 advanced tips share a core goal: turn Claude Code from "a tool you have to watch" into "a tool that works even when you're not there." This article organizes these tips from easiest to hardest and adds key practical details.

Starting from Scratch: Five Steps to Complete Initial Setup

According to the video, setup takes only five steps: download the Claude desktop app (Mac or Windows); log in and choose a paid plan (Pro at $20/month, or $17/month billed annually; Max at $100/month; Ultra at $200/month); click Code at the top of the sidebar; select Local mode below the message box and open your project folder; set the mode to Manual so it asks before every change.

The video specifically clears up a common misconception: although a terminal version exists, everything can be done in the app interface with buttons.

Tip 1: If You Don't Know How, Just Ask Claude Code

According to The Benmore Brief, Claude Code can answer questions about its own features and limitations, and it consults Anthropic's latest documentation. This matters—the video notes that from late October to November, Claude Code released 6 new versions in 10 days, so video tutorials from spring may already be outdated, but the official documentation it consults won't be.

The construction crew lead solved his frustration with exactly this method: he directly asked, "Why do you keep making me approve every little change? Can I turn that off?" Claude Code explained permission modes in plain language and told him the mode selector was right below the message box. One question, and he understood where that evening had gone.

Tip 2: Switch to Auto Mode and Stop Approving Item by Item

Manual mode asks before every edit and every non-read-only command. Anthropic's own documentation points out the problem: by the 10th prompt, you've already started clicking approve without reading; and when you step away, nothing progresses.

According to the video, clicking the mode selector lets you switch: Accept edits mode no longer asks when modifying files, but most other commands still require confirmation; Auto mode skips routine prompts entirely. Before any operation other than reading and writing files, a second AI model checks whether it matches your request and blocks high-risk actions like deleting existing files or sending private data externally. You can also set boundaries in natural language, such as "don't publish anything until I've reviewed it," and Auto will treat it as a stop sign.

The video cautions that Auto reduces prompts, but Anthropic explicitly states it doesn't guarantee safety. Use it for work where you trust the direction.

Claude Code's permission modes are a set of policies controlling "whether user confirmation is needed before a tool call," with common levels including Manual, Accept edits, Auto, and the read-only Plan mode. Anthropic's design logic is: read-only operations like reading files are allowed by default, while operations with side effects—writing files, executing shell commands, network access—trigger confirmation. What's special about Auto mode is that it introduces an independent "review model" for pre-checks; it doesn't look at the full history of the main conversation, only whether the current request matches the proposed action, so it can intercept some unauthorized behavior but may also misjudge or be bypassed. Understanding this mechanism helps you judge when to use Auto and when to fall back to Manual—for production environments, sensitive credentials, or irreversible operations, Manual's item-by-item confirmation remains the last line of defense.

Tip 3: Tell It the Scale You Expect

Anthropic's documentation notes that Claude tends to interpret requests conservatively, building the minimum version that completes the task, because it can't see how big your expectations are. One phrase changes that: "Be ambitious."

The video's example is typical: the materials list initially only had gallons per job. The lead re-requested, "Make the materials list complete, be ambitious," and the result handled cases requiring two coats of paint, leftover paint from last week, jobs with undecided colors, and even automatically added a Monday morning shopping list and a print button. Anthropic says the phrase's purpose is to tell Claude about scope only you know.

Tip 4: Raise the Effort Level for Tricky Parts

According to the video, in the model selector below the message box, effort levels range from low to maximum, with medium as the default. Effort level isn't just thinking time—Anthropic says it controls how much work Claude does on your request: how many files it reads, how much it checks, how far it goes before coming back to you.

The tricky part of the timesheet was overtime calculation: after 7 PM and all day Saturday. The lead set effort to high (Anthropic recommends this for cases where things are easily missed), the system checked more edge cases, and overtime was calculated correctly on the first try. For quick interactions like renaming a button, he set it to low—answers came back faster and used less quota. The maximum level is for the hardest problems; in everyday work it tends to overthink. The video's judgment rule: if it misses steps or doesn't check its work, raise effort; if it clearly tried hard but is still wrong, switch to a stronger model.

Effort level is another dimension in Claude Code distinct from "model selection": it doesn't change the underlying model but adjusts the reasoning budget and tool-call depth. With higher effort, Claude more actively reads relevant files, runs tests, checks edge conditions, and does more self-verification before returning results; lower effort reduces exploration steps, responds faster, and consumes fewer tokens. This aligns with the "extended thinking" approach Anthropic offers in its API, but effort is a coarse-grained knob for interactive scenarios. The key distinction in practice: if the task itself is complex (requiring multi-file coordination, many edge conditions), raise effort; if the task is simple but the model still errs, that's a capability issue—switch to a stronger model rather than continuing to add effort.

Tip 5: Watch the Usage Meter and Let It Finish While You Sleep

Claude Code shares one quota with all your Claude usage. According to the video, the Pro plan has a limit that resets every 5 hours, plus a weekly limit on top. There's a small usage ring below the message box; click it to see how much you've used.

When you hit the 5-hour limit, the pop-up card has a checkbox: "Auto-continue after limit resets." Check it, keep the app open, and go to sleep—after the reset, the system automatically retries the interrupted step. The video emphasizes that this checkbox only appears on the 5-hour card, not the weekly limit card. If you hit the limit every week, Max offers five times the Pro usage.

Tip 6: Use Side Chats for Side Questions

When Claude mentioned it stores work hours in 15-minute increments, the lead wanted to know why. But asking directly in the main chat would interrupt the work in progress. According to the video, press Command+semicolon on Mac, Control+semicolon on Windows, or type /btw in the message box to open a side chat. A side chat can read everything in the current session, but its contents don't go back into the main conversation. Ask, get the answer, close it, and the main work continues as if you never asked. One caveat: side chats aren't saved after you close the app.

Tip 7: Name Your Sessions So They Can Query Each Other

Each session keeps its own conversation, and they quickly pile up in the sidebar. According to the video, click the title at the top of a session to rename it. The lead named his three sessions materials, timesheet, and rain days. The value: Claude Code can read other sessions in the code tab and tell you what they did. So in a new session you can directly ask, "How did we define overtime in the timesheet session?" It will find that session, read it, and answer. It can even pass messages, like "Tell the rain days session that Saturday counts as overtime." By default, it looks at the 20 most recent sessions and skips archived ones.

Tip 8: Review Work with a New Session

The session that built the timesheet knows its own intent, so it reads past its own mistakes. Anthropic's documentation says a second session is the easiest independent reviewer to get, because it never saw the work being written.

The video's approach: click New in the sidebar, hold Option (Alt on Windows) and click the timesheet session so the two sit side by side. Set the new session to plan mode (read-only, no changes), then type "Review the timesheet and materials list, report only places that would cause problems when painters use it, skip style issues." Anthropic warns that a reviewer asked to find problems will always find something, and fixing each one only adds clutter. Among the review results, "employees clocking in after midnight had their hours split across two days" was pasted back into the timesheet session and fixed; another was that the coverage rate for the new primer was unknown, left for later.

Plan mode (read-only mode) is a running state in Claude Code where Claude can read files, search code, and analyze structure but won't execute any writes or commands, making it suitable for code review, design discussions, and risk assessment. Its key difference from Accept edits is "zero side effects": even if the review session makes wrong judgments, it won't contaminate the workspace. Using an independent session for review works like "code review" in software development—the blind spots an author has about their own code are easier to spot from a perspective without contextual baggage. Note that a review session lacks the implicit assumptions made during construction, so it may offer "theoretically correct but practically inapplicable" suggestions; review results still require human judgment on priorities.

Tip 9: Put the Session in Your Pocket

On Thursday the lead was at the job site all day, laptop left in the office. According to the video, before leaving he types /remote control in the materials session, and that session appears under the code section of the Claude mobile app, with a computer icon and green dot. The work still runs on the laptop; the phone is just a window. He can read progress, answer questions, approve requests, and even change modes while on a ladder.

Two things to note: the computer must stay on and the app running; after the laptop sleeps, the session reconnects when it wakes. The video also shows a clever mobile use: photograph the primer can label at the job site, attach it in the app and send "Update the primer with this coverage rate," Claude reads the label, and the materials list in the office updates accordingly.

Tip 10: Put Routine Work on a Schedule

This is the tip promised at the start of the video: let Friday's work start itself. Every Friday the lead drafts next week's schedule from new projects, with fixed steps. He asked in plain language: "Every Friday at 2 PM, draft next week's schedule from new projects in this folder. Schedule outdoor jobs early in the week, keep two-day jobs together, leave the draft for me to review." This created a scheduled task. Click scheduled in the sidebar to see a "Run now" button, a pause toggle, and a history of each run.

The video emphasizes writing instructions assuming no one is there to answer questions, because no one is. Anthropic's advice: describe what success looks like and how results should be handled. Two rules: tasks only run when the app is open and the computer is awake, so he enabled "Keep computer awake" in settings; closing the lid still counts as sleeping. If the computer is asleep at 2 PM, it runs once upon waking. Before trusting it, click "Run now" once to observe.

Claude Code's scheduled tasks are essentially a cron-style scheduler running inside the local app, not a cloud service. This means task reliability depends entirely on the local environment: the app must be running, the computer must not sleep, and the network must be available. Compared with hosted schedulers like GitHub Actions or cloud functions, it wins by directly accessing local files and project context, but sacrifices offline reliability. When setting up, follow the "idempotent + observable" principle: repeated runs shouldn't produce duplicate data, and each run should leave checkable output (like draft files or logs) so you can quickly locate issues during reruns or failures. For critical routine work, keep a manual "Run now" trigger as a fallback.

Where Are the Limits of These Tips?

According to The Benmore Brief's summary, Anthropic is candid about the limitations: Auto mode reduces prompts but doesn't guarantee safety; in very long sessions, the boundaries you draw in chat may be lost; scheduled and mobile features both require the computer to be on and the app open; the cloud version (which runs with the laptop closed) can't see files on your computer; higher effort levels consume more quota; reviewers can always find problems, and you still need to decide what matters.

The video also points out a bigger boundary: when a tool starts holding customer privacy, processing payments, or running all day for many users, someone needs to think about security, backups, and what happens if it breaks on Saturday. Anthropic's official documentation says the defaults are usually right; these tips are for the days when the defaults aren't.

Buying Advice: Pick One Thing to Hand Off First

The video's closing advice is practical: this week, pick one thing you're still sitting and watching Claude Code do, and hand it off using one of these tips. The ten tips can be remembered as a set of actions: ask it how to do things, switch modes, say "be ambitious," match effort levels and watch the usage meter, ask side questions, name sessions, get a second pair of eyes, put it in your pocket, and put routine work on a schedule.

For indie developers and small teams, the common logic of these tips is: move approval from "every step" to "key boundaries," move repetitive labor from "manual trigger" to "scheduled trigger," and move review from "self-check" to "independent session." Claude Code's capability ceiling hasn't changed—what's changed is how and when you step in.

One-line Summary

Ten advanced tips for Claude Code: move from manual approval to automated, remote, and scheduled execution.

Paragraph Summary

Based on analysis from the YouTube channel The Benmore Brief, the article notes that many Claude Code users remain in "beginner mode," sitting at the computer clicking approve item by item, leading to low efficiency and even hitting usage limits. It organizes ten advanced tips from easiest to hardest: asking Claude Code directly, switching to Accept edits or Auto mode to reduce routine approvals, using "Be ambitious" to clarify expected scale, adjusting effort levels by task difficulty, using the usage ring and "auto-continue after limit resets" to let tasks run overnight, using side chats for side questions, naming sessions for cross-session queries, using an independent read-only session for code review, operating the desktop session from your phone via /remote control, and using scheduled tasks to start routine work automatically. The article closes by defining the boundaries of these tips: Auto doesn't guarantee safety, scheduling and remote access depend on the computer staying on, and reviewers can always find problems—and suggests readers hand off one thing this week.