Claude/Claude Code

cbrake · November 24, 2025, 3:08pm

This article makes a good point – just because it uses a particular AI model does not make local tools equivalent. The local agent is responsible for batching up a number of requests to the backend model. How it does this is important.

khem · November 24, 2025, 5:52pm

why > how > what ( in that order of importance )

cbrake · December 31, 2025, 9:00pm

Creating a bookshelf app with Claude Code

cbrake · January 6, 2026, 3:43pm

I’d put the problems in this post at a “good undergraduate” level. They’re accessible to maybe top 5% of CS students at a median US university, and to 85% of CS students at an Ivy. I am not saying Claude Code is a good undergraduate– it’s a different thing altogether. It can do refactors at superhuman speed but can’t publish to npm. What I am saying is that if you’re working at this level of difficulty, Claude Code is a phenomenal coding companion. Setting aside all the productivity arguments, it dramatically amplifies what’s fun about coding by eliminating much of what’s annoying. I would not want to go back to a world without it.

The interesting thing is Slava works at Microsoft …

cbrake · January 19, 2026, 10:13pm

Using Claude Code with proprietary code

When working with proprietary code, you can tell configure Claude not to train on your code:

bradfa · January 20, 2026, 11:06am

Just be careful, as per their consumer terms of service if you supply feedback (ie: filing a bug, using the /feedback command, or answering their “How’s Claude doing today?” thumbs up/down), or if Anthropic silently flag your usage for safety review, then it can still retain your full conversation and use it for training purposes even if you have this setting in your privacy settings turned off: Is my data used for model training? | Anthropic Privacy Center

cbrake · January 20, 2026, 12:58pm

The safety thing is a pretty big hole, and is absent when using the “work” or API products, which probably makes sense if you are doing something really innovative.

OpenCode works pretty well if you want to use the Anthropic API in the terminal. I tried OpenCode with my Claude Max account - it worked for several hours, but then shut me off, as OpenCode somehow simulates a browser to use a Claude Max account, and it appears Anthropic is somehow detecting that. Fortunately, I did not get banned. If you ever use OpenCode with Claude, use the API, not OAuth login.

cbrake · January 23, 2026, 4:25pm

Claude is turning todos into tasks

https://x.com/trq212/status/2014480496013803643

noticed the change already this morning:

cbrake · January 23, 2026, 4:38pm

Regarding security, this looks useful:

cbrake · February 14, 2026, 10:09pm

How I use Claude Code (Meta Staff Engineer Tips)

This video features John Kim, a Staff Software Engineer at Meta, sharing 50 expert tips for using Claude Code (an agentic CLI coding tool by Anthropic) based on six months of daily use. The guide targets engineers moving from traditional IDEs to AI-native workflows, emphasizing context management and validation loops.

Foundations & Setup

Kim emphasizes that proper initialization is critical for Claude Code to understand project architecture.

Project Initialization: Run claude from the project root and use /init immediately. This commands Claude to analyze the codebase and generate a tailored CLAUDE.md file.
Rule Files (CLAUDE.md): This file acts as the “system prompt” or linting rule set for the agent. It operates hierarchically (project-level overrides global) and is read top-to-bottom. Kim suggests keeping it around 300 lines, focusing on high-level architecture, domain context, and build/validation flows.
Keyboard Shortcuts: Efficiency is key. Use Shift+Tab to toggle between “Plan” and “Act” modes, and Double Escape to clear input or rewind context.

Core Workflow Strategy

The video advocates for a “Plan-First” approach to prevent context bloating and errors.

Plan Mode: Always start new features in Plan Mode. Kim argues that rigorous planning and “arguing” with the AI to refine the approach makes the actual code generation trivial.
Context Management: “Context is best served fresh and condensed.” Use /context to visualize token usage and /clear or /compact to manage bloat. Kim warns that bloated context leads to regression in AI performance.
Validation Loops: Define clear validation steps (e.g., “build, run tests, read logs”). Providing Claude with a way to self-verify its work allows it to iterate and fix errors without human intervention.

Power User Features

Kim introduces concepts that treat Claude Code as a composable platform rather than just a chatbot.

Skills & Commands: Recurring workflows (like “fetch Hacker News”) can be saved as Skills, which Claude can then execute via custom slash commands. This effectively creates reusable scripts driven by natural language.
MCPs (Model Context Protocol): While powerful for connecting to external tools (like Figma or Xcode), Kim advises using them sparingly as they significantly increase token usage. He suggests installing only project-specific MCPs.
Sub-Agents: Use sub-agents to handle isolated tasks (like “Architecture Review”). This isolates the context of that specific task, preventing it from polluting the main session’s memory.

Advanced Techniques

Git as a Safety Net: Rely on Git for version control rather than Claude’s internal undo features. Commit CLAUDE.md to the repository to share agent behaviors with the team.
Parallel Workflows: Advanced engineers should run multiple Claude instances simultaneously in split-pane terminals (e.g., iTerm), allowing them to “juggle” context between different tasks while one instance is processing.
YOLO Mode: The --dangerously-skip-permissions flag allows Claude to execute commands without asking for confirmation every time, which is useful for throwaway environments but risky for system-level operations.

bradfa · March 3, 2026, 2:53pm

I wish Claude Code’s planning mechanisms would more clearly show what it has changed between the previous and current iteration of a plan. Especially when plans get longer than just a few paragraphs, it’s often hard for me to understand exactly what’s different in the current iteration of a plan compared to the previous one.

I often end up iterating on a plan many times with the agent. It’s great that Claude Code now puts the plans into files and is able to limit the context usage of iterating on plans better, but having like a diff-version be able to be shown when the agent updates the plan would greatly help me mentally understand how my feedback on a plan changed the plan. Plus it would greatly reduce the time I spend re-reading the same plan when most of it is unchanged or maybe some tiny detail has changed but where I really care about that tiny detail.

cbrake · March 3, 2026, 3:16pm

Could you put something in CLAUDE.md that instructs Claude to store the plan in the git workspace and commit the plan to Git any time it changes the plan? Then you could view plan changes over time like any other file.

In doc-driven dev, I track plans with the source code in the plans/ directory:

github.com/tmpdir-org/tmpdir-claude-code-marketplace

doc-driven-development/commands/plan.md

main

Create a plan file named `plans/<current date>-"$ARGUMENTS".md`. Replace spaces
with '-' in the filename.

If a plan exists in `plans/plans.md` that is marked **IN PROGRESS** then ask the user if they would like to update that plan or start a new one.

Then switch to plan mode and create a plan for coding the current changes and
untracked files (start with `git diff` and `git status`) described in the
README.md and other documentation files. Note, in this phase, we are using
documentation to drive the code, not the other way around. When the plan is
complete, write it to the plan file.

The documentation may be written in past tense as if the feature already exists.
Assume any changes that have not been committed have not been implemented yet,
even though they are described as already implemented. This is a side effect of
documentation driven development.

## Plan Tracking

After the plan is finalized, update `plans/plans.md`:

This file has been truncated. show original

Even then, I have to continually remind Claude I want plans in my repo workspace, not some remote ~/.claude directory.

Why we don’t automatically track plans in Git is a mystery to me. The Plan (and other docs) is the new source code. Why not record the plan and track it like any other source code? Also, it is much easier to review a plan than a boatload of source code, so should not the plan be included in the PR?

bradfa · March 3, 2026, 3:32pm

Yes, in some cases this would be possible, although getting the agent to reliably do that instead of what ever planning functionality Anthropic have baked into Claude Code itself might be hit or miss. Additionally, there’s times when I’m working with Claude on projects where no such infrastructure is present and possibly other contributors to the project are adversarial to AI-use, so tracking plans directly in the project source itself is not realistic unless I change my workflow even more to be able to diligently filter out such git commits.

Ideally, I’d want this to be a native feature of Claude Code. It already stores the plans in a special plans directory, putting some lightweight versioning capability to show me a marked up difference doesn’t seem that burdensome for them to implement.

Maybe it could be directives in CLAUDE.md or a plugin, or some custom workflow, but all of that feels like high friction ways of doing it. But worth considering for sure.

cbrake · March 3, 2026, 10:35pm

Claude is giving me diffs on a plan file I’m working on. Not sure if it would do this if it was stored outside the repo …

cbrake · March 6, 2026, 4:55pm

cbrake · March 14, 2026, 4:44pm

1M context is now generally available for Opus 4.6 and Sonnet 4.6

This seems significant: