I'm Oikon. I typically share Claude Code insights on X (Twitter).
As this eventful year of 2025 draws to a close, I wanted to reflect on Claude Code. For me personally, 2025 was the year AI agents took a massive leap forward. More specifically, it was the year of Claude Code.
The Claude Code CHANGELOG has been updated 176 times total. I have verified and tested every release since at least v1.0.x.
| Version Series | Release Count (as of 2025/12/30) |
|---|---|
| v0.2.x | 37 |
| v1.0.x | 82 |
| v2.0.x | 57 |
| Total | 176 |
Having worked closely with Claude Code throughout the year, I want to organize my thoughts on what I have learned. Please note that these are my personal observations.
TL;DR
Here are the key takeaways from a year of tracking Claude Code:
- Claude Code shipped 176 updates in 2025, demonstrating rapid iteration from beta to v2.0
- Context engineering emerged as the unifying theme across major features
- Key milestones: CLAUDE.md memory files, Plan mode, Subagents, /context command, Skills, and Opus 4.5
- Claude Code's strength lies in its harness design, optimized specifically for Claude models
- Challenges remain: 200k context window (vs competitors' 400k-1M), compaction quality, and knowledge cutoff
- 2026 outlook: Long-running tasks, Swarming capabilities, and enhanced external tool integration
The Shock of First Contact
Before Claude Code
I first touched Claude Code in March 2025, not February when the beta released.
At the time, I was using GitHub Copilot and Cursor while asking ChatGPT and Claude's web version about code, then copy-pasting the good bits and tweaking them locally. Looking back, my development workflow has transformed dramatically over the past year.
Even then, Claude's web version had GitHub integration and could read codebases to answer questions. I personally felt Claude 3.7's coding quality was superior to other AI models at the time.
Meeting Claude Code
In March 2025, I discovered a product called Claude Code. I cannot find my posts from that period since I was not active on X yet, but I remember seeing someone demo it on social media. While IDE-based tools like Cursor and Windsurf were getting attention, the CLI approach caught my eye.
I was deeply impressed when I first tried Claude Code. Previous AI tools worked alongside you while coding, but even the early Claude Code could autonomously code and create multiple components on its own.
Watching an AI agent explore the codebase independently, gather necessary information, and implement features made me realize coding was about to change fundamentally.
From April onward, I continued using Claude Code with pay-as-you-go API billing while getting comfortable with CLI-based coding. Around that time, I started posting on X, praising Claude Code and expressing interest in trying Codex CLI. As a side note, the Codex CLI back then was honestly not very usable, and many people today do not even know it existed. When GPT-5 arrived, Codex CLI transformed dramatically, following a similar trajectory to Claude Code's official release.
"Claude Code is really good, so I want to try Codex too, but the heavy pay-per-use billing makes it hard to use comfortably."
On April 9, the Claude Max plan was announced. On May 1, Claude Code became available at a flat rate through the Max plan. This did not get much attention at the time, but it laid the groundwork for Claude Code's wider adoption.
A few days after that announcement, I subscribed to the Max 5x plan at $100/month. Until then, I had never spent more than $20 on AI tools, so this was a significant shift. Now I use the 20x plan at $200/month, which shows how much things have changed.
"Using Claude Code through Claude Desktop + MCP without API charges is nice, but it still can't match the speed of developing directly with Claude Code. I keep hitting Desktop's chat limits anyway. Better to stop being stingy and just subscribe to the Max plan. +12,000 yen monthly..."
The Official Launch
On May 25, 2025, Anthropic announced Claude 4. I wrote a Zenn article about that moment, which you can read here:
My Zenn article about Claude 4 release
The major changes for Claude Code at this point were:
- Claude 4 (Opus/Sonnet) performance was excellent
- The Max flat-rate plan gained recognition
Claude Code had already shown autonomous capabilities, but with Claude 3.7, the model's performance was not quite sufficient to fully leverage the harness's power. Claude 4's arrival improved how Claude handled tools, reaching a level where it could truly excel.
The Max plan's recognition mattered too. Traditional AI coding involved cautious usage due to pay-per-use API billing. With flat-rate Claude Code access, the AI coding mindset shifted, enabling experimentation and multiple implementation attempts. This flat-rate model has been adopted across the industry, including Codex, Gemini, and some Cursor plans.
Tracking the CHANGELOG
anthropics
/
claude-code
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
Claude Code
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows -- all through natural language commands. Use it in your terminal, IDE, or tag @claude on Github.
Learn more in the official documentation.
Get started
- Install Claude Code:
MacOS/Linux:
curl -fsSL https://claude.ai/install.sh | bash
Homebrew (MacOS):
brew install --cask claude-code
Windows:
irm https://claude.ai/install.ps1 | iex
NPM:
npm install -g @anthropic-ai/claude-code
NOTE: If installing with NPM, you also need to install Node.js 18+
- Navigate to your project directory and run
claude.
Plugins
This repository includes several Claude Code plugins that extend functionality with custom commands and agents. See the plugins directory for detailed documentation on available plugins.
Reporting Bugs
We welcome your feedback. Use the /bug command to report issues directly within…
With Claude 4's release, Claude Code itself received its first major update to v1.0.x. As of December 2025, we are on v2.0.x (v2.0.74 at the time of writing).
I cannot cover all 176 updates, so I will highlight the changes from v1.0.x onward that I found most impactful.
To summarize upfront: many features are strongly tied to context engineering, suggesting that understanding context engineering is essential right now.
Claude Code v1.0.x Era
CLAUDE.md (Memory Files)
CLAUDE.md existed since v0.2.x, but with the v1.0.x release, discussions about what should go in CLAUDE.md intensified. While Cursor Rules existed before, CLAUDE.md became more critical for Claude Code's autonomous implementation.
As the term context engineering spread and the importance of controlling AI agent context windows gained attention, CLAUDE.md's significance grew further.
In August, Codex CLI proposed AGENTS.md, attempting to unify project memory across AI tools. This move reflects the impact CLAUDE.md had on AI coding.
Incidentally, Claude Code does not support AGENTS.md. This has been discussed among users in a GitHub Issue:
GitHub Issue: AGENTS.md Support
Personally, I do not think Anthropic will support AGENTS.md. Here is why:
- Adopting OpenAI's standard poses competitive challenges
- Workarounds already exist in Claude Code
- A universal project memory for all harnesses and models is difficult
Anthropic has created industry standards like MCP and recently Agent Skills, but they seem cautious about adopting others' specifications.
As the GitHub Issue shows, workarounds like @ imports and symbolic links exist. I covered these in my advent calendar article under Day 5:
24 Claude Code Tips - Claude Code Advent Calendar
Finally, I personally find the idea that one AGENTS.md can serve all AI tools somewhat challenging. AI tools combine models and harnesses, and fully leveraging each requires tool-specific project memory. Unlike MCP, AGENTS.md is natural language Markdown, not really a protocol.
This may be why Anthropic is not eager to consolidate around AGENTS.md.
Plan Mode
Plan mode was silently added around v1.0.18 on June 12. You can switch to it with Shift + Tab or set it as the default via /config. If I recall correctly, Cline had introduced Plan mode around April, before Claude Code.
At the time, I was caught up in Vibe Coding and foolishly running --dangerously-skip-permission, so I failed to immediately grasp its value. But as I used Claude Code longer, I came to appreciate Plan mode's utility.
Plan mode matters because it clarifies the AI agent's execution steps beforehand. Many AI coding failures stem from unclear execution sequences, where the agent does not follow the intended implementation path.
Plan mode's value extends to improving downstream bottlenecks. By clarifying specifications upfront, the agent follows them more closely, reducing deviation from expected output. This indirectly reduces review bottlenecks that have become apparent in AI coding. Like Shift Left Testing, eliminating rework early benefits later stages.
When I asked Boris Cherny, a Claude Code developer, about Plan mode's importance, he confirmed it. He said he starts in Plan mode almost every time.
"The single biggest tip is: almost always start in plan mode. Once you align on a detailed plan, Opus 4.5 gets it right pretty much every time."
The latest Claude Code now creates Plan documents in ~/.claude/plans/, moving toward something like Spec Driven Development (SDD). I covered this in Advent Calendar Day 17.
Subagents
Subagents have existed since early Claude Code, operating as the Task tool. I wrote about them in this article:
Starting with v1.0.60, the /agents command enabled custom Subagents.
Subagents have these characteristics:
- Parallel execution capability
- Role specialization
- Independent context windows
The independent context window is particularly important from a context engineering perspective. Current Claude Code context windows are limited, and heavy WebSearch or Read operations quickly fill them.
Subagents help by keeping extraneous information out of the parent context window, offloading it to their own. Given current context engineering constraints, having this escape valve matters.
As AI agent orchestration advances, Subagents' parallel execution capability will become increasingly important.
/context
The /context command was added in v1.0.86 on August 20. Its value lies in visualizing context.
While context engineering as a concept had spread, many people understood it abstractly without truly grasping it.
/context displays:
- System prompt
- System tools
- MCP tools
- Custom agents
- Memory files
- Messages
This makes it immediately clear how much context window is occupied and how much remains.
The /context command helped many users become aware of context windows and understand context engineering's importance. A common example now discussed: MCP tools consume significant space, so removing unnecessary ones helps. This conversation was impossible before /context.
By illuminating the abstract concept of context engineering and making it tangible, the /context command made a major contribution.
Claude Code v2.0.x Era
Opus 4.5
The v2.0.x era's most significant event was Claude Opus 4.5's release on November 24.
At the time, many were migrating from Claude Code to Codex CLI. Reasons included rumors of Claude Code degradation and GPT-5.1 Codex's excellence. I was still primarily on Claude Code with Codex CLI as backup, though Codex had a meaningful share of my tasks.
When Opus 4.5 became available in Claude Code starting v2.0.51, users gradually returned. I say "gradually" because at this stage, AI models had become difficult to distinguish through simple implementation tasks. Many felt they needed hours or days of use to understand the differences. Codex CLI had also evolved into an excellent tool, and Cursor had established its position with Composer-1.
Using it myself, I found Claude Opus 4.5 handled Claude Code's tool specifications well, with satisfactory implementation speed. The affinity between Opus 4.5 and the Claude Code harness felt higher than previous Claude Opus 4.1 or Sonnet 4.5. Claude Opus 4.5 undeniably elevated Claude Code's standing and pushed AI coding history forward.
As an aside, measuring AI model performance has become personally challenging. Simple tasks yield high-quality output from any AI tool now.
SWE-Bench is often cited, but it increasingly feels like marketing performance. Real confidence comes only from hands-on use. A recent YouTube talk by someone from PromptLayer addressed this point engagingly:
Claude Skills (Agent Skills)
Claude Skills was added in v2.0.22 on October 17. On December 18, it became an open standard as Agent Skills.
Skills have been widely discussed lately. I will skip details since many have explained them. This Code Summit video explains Skills well:
You only need to understand the concept of progressive context disclosure for Agent Skills. Context engineering matters here too. Skills enable tool orchestration, and combining that with agent orchestration will gain more attention.
Anthropic emphasizes these Skills characteristics while designing for portability across Claude Code, Claude Desktop, and beyond. As with MCP (Model Context Protocol), Anthropic excels at developing concepts and features essential to AI agents.
Interactive Question Tool
You may not recognize the name "Interactive Question Tool." It is the feature in Plan mode where Claude asks users detailed questions. I adopted this name from the CHANGELOG.
"A video of me trying the newly implemented 'Interactive plan tool' that lets you have a conversation with Claude while planning in Plan mode. Hope the vibe comes through."
I find this feature groundbreaking because it clarifies specification boundaries during planning.
A major cause of AI coding rework was gaps between user-expected specifications and AI implementation. If users could draw clear specification boundaries through prompts or spec documents, problems would not arise. But often those boundaries are ambiguous, causing agents to deviate from user expectations.
The Interactive Question Tool lets the AI agent confirm specification details, much like a human colleague would. This shift from one-way user input to interactive exchange represents major progress in AI tools and significantly impacts AI coding's future.
Plugin Marketplace
Claude Code's Plugin system is another feature I am watching. AI tools are powerful, but many organizations still struggle with team adoption. A major reason is the AI tool learning curve.
Power users of tools like Claude Code can create their own Skills, Hooks, and custom slash commands, but there is a large gap compared to those unfamiliar with AI tools. This creates personalized CLAUDE.md files and individual variations.
The Plugin system distributes functionality, making AI tools more accessible for team development. Having already built an ecosystem for easy installation and sharing, like VSCode Marketplace, is a Claude Code strength.
Creating plugins is simple: just add marketplace.json and plugin.json in a .claude-plugin directory. Here is my demo GitHub project for reference:
oikon48
/
cc-frontend-skills
Claude Code plugin for creating frontend UIs that avoid generic AI aesthetics. (Claude Skills is written in Japanese)
Frontend Skills Plugin
Note: This project is for demonstrating Claude Skills with Claude Code.
A Claude Code plugin for creating distinctive, production-grade frontend UIs that avoid generic "AI slop" aesthetics. This Skill and Subagent is generated with referring Anthropic blog: Improving frontend design through Skills
Features
- Distinctive typography - Effective serif × sans-serif pairings
- Custom color palettes - Beyond default Tailwind colors
- Asymmetric layouts - 35/65 ratios, overlapping elements
- Purposeful animations - Focused on high-impact moments
Demo
- Left: With Skills
- Right: Without Skills
Prompt:
Result:
Installation
Execute following /plugin slash commands in Claude Code.
# Add marketplace
/plugin marketplace add oikon48/cc-frontend-skills
# Install plugin
/plugin install frontend-skills@cc-frontend-skills
Usage
Simply make frontend-related requests and the design skill will be applied automatically:
Build me a flight lookup nextjs web app where the user can put in a flight number and the app gives you the start time, end time, time zones, startWhat Makes Claude Code Excellent
Having reviewed notable CHANGELOG features, I want to address Claude Code's excellence from a tool design perspective. I see many strengths, but I will highlight two.
The CLI Tool Approach
One of Claude Code's strengths is being a CLI tool. CLI tools are usable anywhere by any developer.
Historically, development tools moved from CLI to GUI (IDEs), but in the AI era, rather than returning to CLI, I see it as moving to an Agent UI. In other words:
CLI → GUI → Agent UI
While Claude Code appears to be a CLI tool, I personally feel it pioneered the Agent UI concept.
Claude Code started as Boris's side project, hence its CLI origins. But the CLI design persisted for several reasons, as explained in an Anthropic video:
- Targeting the terminal that developers always use
- LLMs enabled a return from rich UI to chat
- Simple CLI lets users focus on tasks
- CLI is the thinnest wrapper around the model, keeping things clean
The Claude Code Harness
Claude Code's growing prominence in AI coding stems from the harness's completeness.
Throughout this article, I have used "AI model" and "harness." These terms appear in Anthropic's video The future of agentic coding with Claude Code.
AI tools split into AI model and harness, with capabilities emerging from their combination. The video compares the AI model to a horse and Claude Code to a harness (horse equipment), explaining how it controls and guides the AI tool. I love this metaphor and frequently reference it in presentations.
The Claude Code harness is well-designed to appropriately provide tools, prompts, and context to the Claude model, guiding it toward task completion. This differentiates it significantly from other AI tools.
Current Challenges
I have focused on positives, but challenges certainly exist. Many remain. Here are some major ones.
Context Window Size
The first challenge is context window size. GPT-5.2 has 400k tokens, Gemini 3 Pro has 1M, while Claude Opus 4.5 remains at 200k. This significantly constrains users.
Claude Code mitigates context window issues through Subagents and recently implemented rules, but even the harness's power cannot fully hide this problem.
Fortunately, I had access to Claude Sonnet 4.5 [1M] and used it primarily before Opus 4.5 arrived. I remember feeling how much easier things became with a larger context window. Conversely, imagine what happens when Claude gets a standard 1M context window.
Compaction Quality
Even accepting Claude's current 200k context window, Claude Code's context compaction (/compact, auto-compact) has room for improvement.
I have felt Codex CLI handles compaction better in some situations. Claude Code seems to simply summarize the session unless you explicitly specify compaction instructions.
If model context windows grow larger, compaction may become less relevant. But for now, you either need compaction instructions or frequent /clear commands to keep the context window clean.
Knowledge Cutoff
This universal model problem affects Claude and others: knowledge cutoff impact appears frequently. Specifically, I often encounter bugs from models not knowing the latest libraries or versions.
Workarounds include using WebSearch or Context7 MCP to fetch current information, but ideally the Claude Code harness would handle this automatically. As Simon Willison's recent article suggested, centralizing version information and loading it via Skills when needed seems like a good approach.
Claude Code is still a product barely ten months old. I look forward to watching it grow.
Should You Use Third-Party Tools?
Claude Code's user base grew dramatically this year, and third-party tools proliferated. Notable examples include awesome-claude series collecting custom slash commands and Subagents (various authors), plus complex harness mechanisms built with Claude Hooks.
My personal opinion: use these third-party tools cautiously.
Claude Code works through the combination of the Claude model and Claude Code harness. Many third-party tools essentially build additional harness layers outside the existing harness.
Claude model and tool evolution is dramatic, and Claude Code's behavior changes significantly week to week. Unless third-party tools are maintained to keep pace, many become unnecessary harness overhead.
Given that future Claude Code versions may incorporate and improve upon these features, I personally do not recommend heavy third-party tool dependence.
The Future of Claude Code
In just ten months of 2025, Claude Code evolved dramatically. Having shared my personal views on Claude Code's present, I want to speculate about 2026.
Evolution Trajectories
Claude Code has established its position as an AI coding tool, but significant evolution potential remains. Drawing from hints gathered when I asked Claude Code developer Boris at the recent Claude Code Meetup Tokyo, I will explore possible directions.
"I asked Boris Cherny about Claude Code's future direction at the Claude Code Meetup Tokyo Q&A. I was curious about orchestration and scheduling. His answer: He is very interested in two things—'Long running' and 'Swarm'. Both seem to be at the demo stage. Quite a hot topic."
Long-Running Tasks
Claude Code already acts quite autonomously, but not completely.
For tasks like reviewing an entire codebase and performing certain operations, Claude Code sometimes stops before completion. Users must then say "Continue, you're not done yet" to resume.
Future improvements should enable Claude Code to run longer and complete tasks.
Notably, an official Anthropic plugin called ralph-wiggum already enables long-running execution. This may become standard or improved.
"The ralph-wiggum plugin is a mechanism that creates a 'User prompt submit hook' to automatically inject '/ralph-loop' as user input. While in ralph-loop, it infinitely injects '/continue' as user input when Claude's turn ends."
Swarming (Orchestrated Execution)
Orchestrated execution was also mentioned. This differs from parallel Subagent execution.
Around June 2025, using tmux to run multiple Claude Code instances with Manager and Worker roles for orchestration became a topic of discussion.
Claude Code may eventually support Swarming natively. I personally suspect the undocumented delegate mode may be part of swarming functionality.
"Claude Code 2.0.71 quietly added a mysterious 'delegate mode'"
External Tool Integration
This is my personal prediction, not from Boris.
Currently, Claude Code has standard tools plus Skills and MCP tools that it uses autonomously. However, there is room for improvement in selecting the right tool at the right time.
MCP tools in particular are criticized for consuming context windows, suggesting tool call mechanisms need enhancement. This topic came up at Anthropic Builder Summit Tokyo, with discussions about using programs to mediate MCP tool calls.
Regarding MCP, a beta feature for dynamically loading MCP tools has been implemented and is available for testing.
"I saw a post saying that the feature to dynamically load MCP tools in Claude Code is available as a beta release, so I tried it. Setting these two environment variables: ENABLE_TOOL_SEARCH=true and ENABLE_EXPERIMENTAL_MCP_CLI=false lets you try it out. When I tested it, MCP tool context became 0 and I confirmed tools are loaded dynamically when used (see next post). This will likely be included in a Claude Code release soon."
Expanding further on external tools: Physical AI will likely gain more attention in 2026. Robots and autonomous vehicles are already exciting areas, and I expect Claude Code to eventually perceive and interact with the physical world through external sensors.
The Value of Following Claude Code
Claude Code is an excellent tool. Since Opus 4.5's release especially, it has pulled ahead. But other AI tools have also improved significantly, and honestly, users should pick whatever suits their use case and preferences.
Even in this era of competing AI tools, Claude Code (and Anthropic) consistently sets milestones at the frontier, making it personally worthwhile to use and follow. At minimum, Claude Code led AI tools in 2025. The features and mechanisms Claude Code released will undoubtedly be recorded when we look back on engineering history years from now.
For those planning to work in engineering for the next decade or more, Claude Code is worth following.
Conclusion
I wrote this article to organize my personal thoughts from watching Claude Code throughout the year. I deeply realize how impactful this year was for me personally. I am genuinely grateful to be an active engineer during such fascinating times.
Next year feels like it will accelerate even more, but I intend to keep enjoying myself while occasionally gritting my teeth to avoid being left behind.
Side note: I was inspired to write this by the following article. It is long but excellent, so please take the time to read it in full:
Sankalp's Experience with Claude Code 2.0
If you found this article valuable, follow the author on X (@oikon48) for ongoing Claude Code insights and updates.
References
GitHub
anthropics
/
claude-code
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
Claude Code
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows -- all through natural language commands. Use it in your terminal, IDE, or tag @claude on Github.
Learn more in the official documentation.
Get started
- Install Claude Code:
MacOS/Linux:
curl -fsSL https://claude.ai/install.sh | bash
Homebrew (MacOS):
brew install --cask claude-code
Windows:
irm https://claude.ai/install.ps1 | iex
NPM:
npm install -g @anthropic-ai/claude-code
NOTE: If installing with NPM, you also need to install Node.js 18+
- Navigate to your project directory and run
claude.
Plugins
This repository includes several Claude Code plugins that extend functionality with custom commands and agents. See the plugins directory for detailed documentation on available plugins.
Reporting Bugs
We welcome your feedback. Use the /bug command to report issues directly within…
GitHub Issue: AGENTS.md Support
agentskills
/
agentskills
Specification and documentation for Agent Skills
Agent Skills
Agent Skills are a simple, open format for giving agents new capabilities and expertise.
Skills are folders of instructions, scripts, and resources that agents can discover and use to perform better at specific tasks. Write once, use everywhere.
Getting Started
- Documentation - Guides and tutorials
- Specification - Format details
- Example Skills - See what's possible
This repo contains the specification, documentation, and reference SDK. Also see a list of example skills here.
About
Agent Skills is an open format maintained by Anthropic and open to contributions from the community.
anthropics
/
claude-plugins-official
Anthropic-managed directory of high quality Claude Code Plugins.
Claude Code Plugins Directory
A curated directory of high-quality plugins for Claude Code.
⚠️ Important: Make sure you trust a plugin before installing, updating, or using it. Anthropic does not control what MCP servers, files, or other software are included in plugins and cannot verify that they will work as intended or that they won't change. See each plugin's homepage for more information.
Structure
-
/plugins- Internal plugins developed and maintained by Anthropic -
/external_plugins- Third-party plugins from partners and the community
Installation
Plugins can be installed directly from this marketplace via Claude Code's plugin system.
To install, run /plugin install {plugin-name}@claude-plugin-directory
or browse for the plugin in /plugin > Discover
Contributing
Internal Plugins
Internal plugins are developed by Anthropic team members. See /plugins/example-plugin for a reference implementation.
External Plugins
Third-party partners can submit plugins for inclusion in the marketplace. External plugins must meet quality and security standards for approval.
External Articles
Simon Willison - actions-latest
Sankalp's Experience with Claude Code 2.0
Videos
Top comments (0)