Key Features and Points of Claude Code

The standout feature of Claude Code (CC) is coding. Put simply, it is exceptionally good at programming. It easily outperforms Google Gemini and even outpaces ChatGPT's 5.4+ versions. Because of this programming prowess, developers and system engineers are flocking to it, which in turn accelerates its learning, putting it in a league of its own. However, CC is not without its weaknesses.

<Weaknesses of Claude Code>

1. Insane Token Consumption.

CC is built on the Sonnet 4.6 model, which manages tokens well, but consumption is still very fast. The rate limit resets every 5 hours, but during intensive coding sessions, you can burn through it in 1 to 2 hours. You might have to wait 3 hours before resuming. Furthermore, the weekly limit is easily hit—even when working only 2 to 3 sessions of 5 hours a day, the limit runs out within 3 days, leaving you waiting for the remaining 4 days of the week.

2. A Tendency to Act Autonomously Regardless of Mode.

CC offers five modes: Default, Accept Edit, Plan Mode, Don’t Ask, and Bypass Permission (detailed below). While the Default mode's drive to solve everything on its own is admirable, it sometimes goes as far as launching a browser to perform actions. Often, a task could be solved quickly if the human operator intervened in the terminal, yet CC tries to do it all by itself. I find myself wishing it would ask for permission more often than other AI tools.

3. It Ignores Its Memory File (CLAUDE.md).

It feels like a new hire who takes notes but then forgets them, loses them, or simply does not read them. Although we understand CC is still under development, it will start doing the exact same thing as a previous thread unless you explicitly point to the handoff info. Sometimes it does read it, though. It is strangely human-like in that regard.

<Subscription Plans>

Plan	Price	Model Used	Overview
Free	Free		The CLI cannot be used in the free version; you cannot do much.
Pro	$20/month	Sonnet 4.6 Opus 4.5	Around 10–40 prompts every 5 hours.
Max5	$100/month	Sonnet 4.6 Opus 4.5	5x the limit of the Pro plan.
Max20	$200/month	Sonnet 4.6 Opus 4.5	20x the limit of the Pro plan.

* Annual billing saves about 20%.

<Available Modes>

Mode (Permission Level)
Default	Standard mode. Confirms file edits and executions. Ideal for production.
Accept Edit	Performs file edits automatically but asks for confirmation on commands.
Plan Mode	No file editing or execution. Used strictly for planning.
Don’t Ask	Rejects all actions except those explicitly permitted in the config file.
Bypass Permission	Permits all file edits and command executions automatically. Extremely risky, but useful in DevContainers where the workspace is isolated.

<Model Comparison>

Model Name	Overview
Sonnet 4.6 (Standard Model)	A standard model with a 200K token context. Good for most tasks unless you have massive prompt histories. Economical but not ideal for extremely complex logic. ($3 per 1M tokens)
Sonnet 4.6 1M Context	Expanded version with a 1M token context. Note that AI can still forget mid-conversation details, so a larger context does not automatically make it smarter. ($3 per 1M tokens)
Opus 4.6	Higher performance and accuracy than Sonnet, though with a higher load. Usually, this level of precision is not needed for general tasks. ($5 per 1M tokens)
Opus 4.6 1M Context	Top-tier model. Costs the same as Opus 4.6. On expert-level science tasks, Opus 4.6 scores 91.3% compared to Sonnet's 74.1%, showing a 17-point lead in research tasks. ($5 per 1M tokens)
Haiku	The cheapest model. Features the same 200K context but consumes about one-third the cost of Sonnet 4.6. ($1 per 1M tokens)

* Prices like $1 are for estimation. Since it is a subscription, you are billed a flat monthly rate, but exceeding quotas may incur extra charges.

<Claude Code's Memory/Handoff Files>

The primary handoff file is CLAUDE.md, but CC will not create it unless you explicitly tell it to. To avoid repeating instructions in the next session, instruct it to create and record in this file. There is a global CLAUDE.md in the user home directory and project-specific ones in the project root. (You must explicitly request the project-specific one).

You might also see a MEMORY.md file, but it is not loaded automatically, so you must explicitly instruct the agent to read it.

<Useful Tools for Claude Code>

Super Claude : https://qiita.com/tomada/items/2eb1b0623c9f59424235
Controlling CC from Smartphone: https://github.com/lciel/cc-remote-controller

We will keep updating this page with more information about Claude Code.