Claude Code review: what it is good at and where it wastes your time

Filed under Review

By Gerald · 17 August 2026

Lines of syntax-highlighted code filling a dark computer screen

I build Flow with Claude Code running most days, and I have stopped assuming it will handle a task well just because the task is "coding." Some weeks it clears a backlog I would have put off for a month. Other weeks I watch it circle a problem that needed five minutes of my own judgment, not another prompt.

The short answer: Claude Code is worth running as your default terminal agent if your work has a lot of mechanical, well-specified tasks, and it is a poor fit for anything that depends on taste or a fuzzy brief. I am testing version 2.1.220, current as of late July 2026.

A terminal agent is a force multiplier on work you already know how to describe. It is not a replacement for deciding what to build.

What Claude Code is, in one paragraph

Claude Code is Anthropic's agentic coding tool. It runs in your terminal (also as a VS Code and JetBrains extension, a desktop app, and a browser session at claude.ai/code), reads your codebase, edits files across multiple locations, runs shell commands, and opens pull requests. You describe a task in plain language and it plans an approach, works through the files involved, and verifies its own output where it can. It is not an autocomplete layer sitting inside an editor. It is closer to a junior engineer you brief in a sentence and then either supervise closely or let run.

The tasks it handles better than I expected

A cluttered workshop bench with hand tools, clamps, and a drill press — Delegating mechanical, well-specified work is the clearest win, whether the bench is physical or a codebase.

Large mechanical refactors are the clearest win. Renaming a field across twenty files, updating a dependency and fixing the breakage that follows, rewriting a batch of components to use a new hook. These are tasks with a correct answer that a human would find tedious to execute by hand. Claude Code grinds through them faster than I would, and it does not get sloppy on file eighteen the way I might.

Test writing and test fixing come next. Point it at an untested module and ask for coverage, and it writes real tests, runs them, and iterates on failures without me watching every step. Same with a red CI run: paste the failure, and it traces the stack, finds the change that broke it, and proposes a fix.

Git and PR mechanics are a genuine time saver. It stages changes, writes a commit message that actually describes what changed, and opens a pull request with a body worth reading. I still write my own commit messages for anything sensitive, but for routine work I let it draft one and edit lightly.

Cross-file exploration in a codebase I do not remember well is where it earns its keep fastest. "Where does this setting get read, and what happens if I change the default" is a question that used to cost me ten minutes of grepping. Now it costs one prompt.

The tasks where I stopped delegating

Anything with a fuzzy spec goes badly. If I cannot describe the outcome precisely, Claude Code cannot either, and it will confidently produce something plausible that misses the actual intent. The fix is not a smarter model. It is a clearer brief, which is work I have to do myself before I can hand anything off.

Visual and interaction design decisions are a bad fit. It can implement a spacing change once you tell it the exact value, but it cannot tell you whether a layout feels right. I still make every real design call by eye, in the browser, myself.

Architectural decisions that trade off long-term maintainability against short-term speed need a human who will live with the consequences. Claude Code will give you a reasonable option if you ask, but it has no stake in which one you pick, and it will implement whichever one you choose with equal confidence.

The tasks that quietly cost more than doing them yourself are small, ambiguous edits scattered across a codebase you know well. If you can hold the whole change in your head and type it faster than you can explain it, explaining it is the more expensive path. I catch myself doing this less often now, but it still happens on a rushed day.

What it costs in practice

Claude Code is included in the Pro plan (17 dollars a month billed annually, 20 dollars billed monthly), both Max tiers (from 100 dollars a month), Team (Standard seat 20 dollars a month annually or 25 dollars monthly, Premium seat 100 dollars annually or 125 dollars monthly), and Enterprise, where you pay a per-seat minimum plus usage at API rates. The free plan does not include it. These figures are from Claude's official pricing page as of July 2026.

If you connect it to a Claude API key instead of a subscription, you pay per token at standard API rates, and Anthropic's own cost guidance for enterprise deployments puts the average around 13 dollars per developer per active day, or 150 to 250 dollars per developer per month, with spend staying under 30 dollars a day for 90 percent of users. Heavy sessions with a large codebase, Opus as the default model, or multiple parallel agent instances push that number up fast. Running an agent team of subagents uses roughly seven times the tokens of a single session, because each teammate keeps its own context window running until it finishes.

The habits that keep cost sane: clear the conversation between unrelated tasks instead of letting context pile up, use Sonnet rather than Opus for routine work, and disable MCP servers you are not actively using in a given session, since idle tool definitions still occupy context space.

Permissions, sandboxing, and the things it should not touch

Claude Code uses a tiered permission system. Read-only actions like file reads and searches run without asking, restricted to your working directory and any folders you have explicitly added. Bash commands require approval every time, except for a built-in allowlist of safe read-only commands, and an approval can be remembered permanently per repository and command. File edits require approval each time by default, and that approval lasts until the session ends.

Recent releases add finer control on the network side: a sandbox.network.strictAllowlist setting denies any host not explicitly allowed for sandboxed commands, without prompting you each time. That matters if you run Claude Code against a codebase with real credentials or production access nearby, because the default posture is permissive once you have granted an approval, not paranoid by default.

My own rule: I never let it run in a mode that skips permission prompts on a machine with production credentials configured. I use plan mode for anything I have not fully specified, so it proposes an approach before touching a single file, and I keep an eye on what it stages before I let a commit go through. It has never done anything malicious to my codebase. It has occasionally done something technically correct that I did not actually want, which is the more common failure mode worth guarding against.

Who should use it and who should stay in an editor

Claude Code suits people with real backlogs of mechanical work: test coverage gaps, dependency upgrades, repetitive refactors, CI maintenance. It also suits anyone comfortable specifying a task in writing before handing it off, because the quality of the brief is the ceiling on the quality of the result.

It is a weaker fit if most of your day is small, exploratory edits where typing the change is faster than describing it, or if your work leans heavily on visual judgment that a terminal session cannot see. Plenty of developers run both: an AI editor like Cursor for the fast, in-context edits, and Claude Code for anything large enough to delegate outright. I do exactly that. Read my Cursor review if you want the other half of that comparison, and Claude Code compared with Cursor if you want the two side by side directly.

I built Flow with a mix of both tools plus a lot of manual work in between. Claude Code is genuinely good at the parts of building a SaaS product that are pure execution once the decision is made. It has never once told me what to build. That part stays mine, and I think it should stay yours too.

Frequently asked questions

How much does Claude Code cost? It is included in Claude's Pro plan (from 17 dollars a month billed annually), both Max plans (from 100 dollars a month), Team plans (from 20 dollars a seat annually), and Enterprise. Used with an API key instead of a subscription, you pay per token, with enterprise deployments averaging around 13 dollars per developer per active day.

Is Claude Code better than Cursor? They solve different problems. Cursor is an editor with AI woven into every keystroke, best for fast, in-context edits you are watching happen. Claude Code is a terminal agent you brief and then let run, best for larger, well-specified tasks you can fully delegate. I use both, for different halves of the same day.

Does Claude Code work with any programming language? Yes. It reads and edits any text-based codebase, so language support depends on the model's general coding ability rather than a fixed list. It has handled everything I have thrown at it across a TypeScript and Convex stack without needing language-specific configuration.

Can Claude Code run commands on my machine without asking? Not by default. Bash commands need your approval every time except for a small allowlist of safe read-only commands, and file edits need approval until you end the session. There are modes that reduce prompting for trusted workflows, but the default posture asks before it acts, and network access can be locked to an allowlist with a sandbox setting.

Is a terminal agent worth it if I already use an AI editor? For most developers, yes, because the two tools are strong at different things. An editor's AI is optimized for the file you have open right now. A terminal agent is optimized for a task described in full, run across as many files as it takes, while you do something else. If your work has enough delegatable tasks to justify the context switch, running both pays off.

My verdict

Delegate the mechanical work and keep the judgment calls. Claude Code is worth the subscription the moment your backlog has more than a few tasks you can describe fully in a sentence. It is not worth reaching for on the tasks where you would need three sentences to explain what "correct" even means, because at that point you are doing the hard part yourself anyway.