In recent versions, Claude Code has changed how thinking mode works compared to the version in Claude Code in Action.

From https://github.com/anthropics/claude-code/issues/9072#issuecomment-3648376741:

• Only ultrathink is the special keyword that enables thinking on a per-request basis. Phrases like “think”, “think hard”, “think more” don’t have any impact on the allocated thinking token budget.
• If you need fine-grained control over the token budget (vs. just allocating all 31,999 tokens to the thinking budget), you can set the MAX_THINKING_TOKENS environment variable. This setting takes priority over ultrathink, so if you set a lower MAX_THINKING_TOKENS threshold, ultrathink doesn’t override it.

From https://code.claude.com/docs/en/common-workflows#use-extended-thinking-thinking-mode:

• Sonnet 4.5 and Opus 4.5 have thinking enabled by default. All other models have thinking disabled by default.
• When thinking is enabled (via /config or ultrathink), Claude can use up to 31,999 tokens from your output budget for internal reasoning.
• Note that ultrathink both allocates the thinking budget AND semantically signals to Claude to reason more thoroughly, which may result in deeper thinking than necessary for your task.

• #claude

A project using devenv prompts: do you want to allow configuration setting 'extra-substituters' to be set to 'https://devenv.cachix.org' (y/N)?., but I cannot input y in the stdin. It happens in fish shell, and there is an issue: https://github.com/direnv/direnv/issues/1022.

The solution is to set accept-flake-config: true in the global nix setting (via):

{
  nix = {
    settings = {
      experimental-features = "nix-command flakes";
      accept-flake-config = true;
    };
  };
}

• #bib
• #devenv
• #nix

The following configuration enables Emacs to inherit the correct $PATH from the shell in all cases:

• M-x describe-variable: exec-path
• M-x eval-expression: (getenv "PATH")
• M-x shell-command: echo $PATH
• M-x compile: echo $PATH

My shell is located at /run/current-system/sw/bin/fish, so the configuration is:

(use-package exec-path-from-shell
  :pin nongnu
  :if (memq window-system '(mac ns x))
  :init
  (setq shell-file-name "/run/current-system/sw/bin/fish")
  :config
  (dolist (var '("__fish_nixos_env_preinit_sourced"
                 "__NIX_DARWIN_SET_ENVIRONMENT_DONE"
                 "__HM_SESS_VARS_SOURCED"))
    (add-to-list 'exec-path-from-shell-variables var))
  (exec-path-from-shell-initialize))

(use-package envrc
  :pin melpa
  :hook (after-init . envrc-global-mode))

• shell-file-name specifies the shell for commands like shell-command and start-process, it uses your system’s default shell by default.
• exec-path-from-shell-shell-name specifies the shell that exec-path-from-shell uses to retrieve the $PATH environment variable, it falls back to shell-file-name.
• exec-path-from-shell-variables is recommended to set if you use Direnv + HomeManager + Fish (via).

• #emacs

I was installing claude-code via Nix, but new releases take 2-3 days to appear in the nixpkgs-unstable branch. I want to use npm directly instead. By default, npm installs global packages to the immutable Nix store, causing npm install -g fails.

Based on this article, the solution is straightforward: configure npm to use a writable prefix directory. The required Nix configuration is (commit):

home.file = {
  ".npmrc".text = "prefix=~/.npm-global";
};

home.sessionPath = [
  "$HOME/.npm-global/bin"
];

Now I can install claude-code using npm:

npm install -g @anthropic-ai/claude-code

The version today is:

$ claude --version
1.0.72 (Claude Code)

• #nix
• #npm

https://anthropic.skilljar.com/claude-code-in-action

• Tips

  • #: set memory
  • @: mention files
  • Ctrl-v: paste image

• Context

  • Escape: interrupt and redirect the conversation
  • Double-tap Escape: rewind the conversation
  • /compact: summarize current conversation for things learned and continue a related task
  • /clear: dump current conversation and start an unrelated task

• Intelligence

  • Planning mode for breadth-first tasks
  • Thinking mode for depth-first tasks. Use keywords (sorted from less to more tokens):
    • Think: Basic reasoning
    • Think more: Extended reasoning
    • Think a lot: Comprehensive reasoning
    • Think longer: Extended time reasoning
    • Ultrathink: Maximum reasoning capability

• Custom commands

  • Create a markdown in .claude/commands/<COMMAND>.md.
  • In the command, use $ARGUMENTS to receive user prompt. For example,
    • Create .claude/commands/test.md:

      Your task is to write tests for $ARGUMENTS
      

    • Usage: /test the User model
  • I found if you create commands in ~/.claude/commands/, they become global. For example,
    • Create ~/.claude/commands/backup.md:

      Create a backup for ~/.claude/CLAUDE.md using gist
      

    • Use in any conversation: /backup

• MCP servers

  • Add server

    claude mcp add playwright npx @playwright/mcp@latest
    

  • Pre-configure permissions: open the .claude/settings.local.json file and add the server to the allow array:

    {
      "permissions": {
        "allow": ["mcp__playwright"],
        "deny": []
      }
    }
    

• #claude
• #course

I’m using the Atlassian Remote MCP Server in Claude Code to fetch Jira tickets. I installed it per-project using claude mcp add-json <name> <string>. Recently, I noticed it keeps failing in one project, although the config is exactly the same.

I checked the problem by locating the MCP server command and executing it within the project. The error raised: No environment file found at: `.env`. It reminded me that UV_ENV_FILE=.env is set globally, and it’s too convenient to be noticed. After removing the environment variable, the MCP server works.

• #python
• #uv

When using Nix to manage my development environment, Claude’s MCP servers couldn’t find tools installed via Nix. The issue was missing Nix paths in the environment.

I found Debugging Claude’s MCP Server with Nix which describes a method to debug MCP server environments by dumping PATH and environment variables:

{
  "mcpServers": {
    "debug": {
      "command": "/bin/sh",
      "args": [
        "-c",
        "echo $PATH > /tmp/claude_path.txt; env > /tmp/claude_env.txt"
      ]
    }
  }
}

By comparing the output files, I discovered /etc/profiles/per-user/james/bin was missing from PATH. The fix is to explicitly prepend it to PATH:

{
  "mcpServers": {
    "github": {
      "command": "/bin/sh",
      "args": [
        "-c",
        "PATH=/etc/profiles/per-user/james/bin:$PATH exec npx -y @modelcontextprotocol/server-github"
      ],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "<token>"
      }
    }
  }
}

• #bib
• #mcp
• #nix

Install llm:

uv tool install llm

Install plugins:

llm install llm-mlx llm-ollama llm-gemini llm-mistral llm-openrouter llm-sentence-transformers llm-cmd llm-jq

Install llm with its plugins in one line (great to re-install the environment):

uv tool install llm --with llm-mlx --with llm-ollama --with llm-gemini --with llm-mistral --with llm-openrouter --with llm-sentence-transformers --with llm-cmd --with llm-jq

llm-mlx requires Python 3.12 or lower, there are two ways to set Python version.

Environment variable:

export UV_PYTHON=3.12

Command line option:

uv tool install llm --python 3.12

Update tools is as easy as uv tool upgrade --all.

• #llm
• #python

Hugo-modus supports multilingual now, it’ll display translation links for multilingual posts. As I’m blogging using Denote and Hugo, I changed the #+export_file_name by appending the locale such as: post-1.zh, expecting it to export post-1.zh.md but failed. The complete configuration is as the following:

#+hugo_base_dir: ~/code/hugo-modus/exampleSite
#+hugo_section: posts
#+export_file_name: post-1.zh

Then I found there is already a solution, which is as easy as appending .md, such as #+export_file_name: post-1.zh.md. The result is on the example site.

• #emacs
• #hugo

shot-scraper is the second tool created by Simon Willison that I’ve installed on my machine, the first being llm.

I install Python tools using uv.

uv tool install shot-scraper
shot-scraper install # install browsers

Take a screenshot:

shot-scraper shot https://yejun.dev

Record an HTTP Archive (HAR) file:

shot-scraper har https://yejun.dev

View the file using Chrome HAR Viewer.

• #python
• #scraping

This blog is built with Hugo, and the content is exported from Org files to Hugo-compatible Markdown via ox-hugo. Each heading is followed by an anchor after the export:

## Heading {#heading}
### Another heading {#another-heading}

To make the anchors clickable, Kaushal Modi (the author of ox-hugo) recommends replacing anchors with links (source), the steps are as follows:

First, create a partial layouts/partials/headline-hash.html:

{{ . | replaceRE `(<h[2-9] id="([^"]+)".+)(</h[2-9]+>)` `${1}&nbsp;<a class="headline-hash opacity-0" href="#${2}">#</a> ${3}` | safeHTML }}

Next, apply the partial to .Content:

{{ partial "headline-hash.html" .Content }}

The result will look like this:

<h2 id="heading">Heading&nbsp;<a class="headline-hash" href="#heading">#</a> </h2>
<h3 id="another-heading">Another heading&nbsp;<a class="headline-hash" href="#another-heading">#</a> </h3>

• #emacs
• #hugo

• Lock/Unlock Win: Fn + F9
• Lock/Unlock Function keys: Ctrl + Fn

• #keyboard
• #cherry

By default, multiple calls of the content_for helper using the same identifier will concatenate and output them together.

Set flush: true to flush all previous content_for calls:

<% content_for :example %>
This block is flushed.
<% end %>

<% content_for :example flush: true do %>
This block will display.
<% end %>

• #rails