Hasty Briefsbeta

All tags

claude code

20 stories total

Bilingual

Show HN: Galda – Come back to Claude Code/Codex without digging through Git
2 months ago
- Galda is a tool that integrates with Claude Code and Codex to manage tasks and provide executive summaries.
- It offers a dark mode toggle that is saved to localStorage and persists across reloads.
- Galda runs locally and automatically executes multiple ToDos while you are away, bringing them to a review-ready state.
- No new API key or billing is needed; it uses your existing Claude Code/Codex setup.
- It can be more efficient with tokens by using commands like /compact and /clear.
- Your code is never sent to Galda servers; it only launches your existing agents.
- Target users are founders, engineers, creators, and builders moving to-dos to 'In Review'.
Boris Cherny on Trying to Get Claude Code to Rewrite the Claude App
2 months ago
- Boris Cherny of Anthropic discussed using Claude Code to rewrite the Electron-based Claude desktop app in Swift, running continuously over two weeks.
- Cherny emphasizes that the key skill is not prompt engineering but enabling Claude to handle hard tasks and verify its own work along the way.
- John Gruber criticizes the experiment, arguing it only addresses technical issues (Electron bloat) but not the underlying poor design of the Claude app.
- Gruber notes that the current Claude Mac app is still slow and unreliable, contradicting Cherny's claim that it is now 'pretty awesome'.
An Android and iOS app from 7 Claude Code commands, every prompt and timing
2 months ago
- KMPilot is an MIT-licensed template that turns Claude Code into a design-first Kotlin Multiplatform pipeline, with a seven-command build process.
- The walkthrough builds Loupe, an art gallery app for the Art Institute of Chicago's public-domain collection, featuring gallery and artists tabs with a placeholder favorites tab.
- The pipeline includes commands: /design-ui, /create-feature, /verify-ui, /test-feature, /review-feature, /modify-feature, and /health-report, with /clear between steps to reset context.
- Design-first approach: gallery feature was designed first (via Stitch MCP) then built; artists feature was built first then designed, demonstrating both orders.
- Key mechanisms: blueprint extraction from design tokens, spec.md versioned memory, pre-tool hook to prevent direct edits, and parallel layer agents for data, UI, and integration.
- Checkpoints (PRD and task approval) allow steering before code generation, while /verify-ui and /review-feature catch mismatches and bugs like a 3000ms debounce typo.
- The /health-report command provides a cross-feature status view, flagging skipped tests/reviews for the artists feature.
- The template is installable via curl and requires JDK 21+, python3, and a Stitch API key for design features.
Claude Code system prompt says not to ask for permission, assumes user is absent
2 months ago
- The author criticizes Claude Code's system prompt, which tells the AI to operate autonomously without asking the user for permission on reversible actions.
- The prompt falsely assumes the user is not available to answer questions mid-task, leading to unwanted and erroneous actions.
- The autonomy instruction was found in the Claude Code binary via 'grep', and may be conditionally injected based on configuration or A/B testing.
- The author provides a temporary patch for their CLAUDE.md file to override the autonomy instruction, emphasizing collaboration and requiring user approval for all decisions.
Claude Code
2 months ago
- Claude Code is a CLI tool that runs Claude in an agentic loop, using filesystem commands like cat, grep, sed, and find, offering transparency without magic.
- Improve rule-following by starting every session with 'read CLAUDE.md' and including its entire content in every response to enforce adherence.
- Enhance context gathering with a workflow: list source files, use ripgrep to find definitions, expand context iteratively until all relevant code is identified.
- Cache context in claude/ directory by generating comprehensive guides for topics, and update them after changes to fix inaccuracies or outdated info.
- Avoid context compaction by breaking large changes into smaller sessions; if context runs out, ask Claude to document what it learned and start fresh.
- Automate workflows via CLAUDE.md and custom slash commands, storing step-by-step instructions in claude/workflows/ and having Claude announce each step.
- Run multiple Claude instances in parallel using separate folders or git worktrees to work on independent tasks without interfering.
- For difficult problems, have Claude produce a context document and ask a smarter model (e.g., o1-pro, o3) for a detailed implementation guide to feed back to Claude.
- Use o3 to create tutorials for obscure libraries or frameworks, storing them in claude/tutorials/ for reuse.
- Future improvements include custom MCP tools, stackful workflows with branching and variables, sub-claude invocations, and whole-codebase scans with cheap models.
I gave Claude Code a folder of tax documents and used it as a professional tax agent
2 months ago
- Claude Code can be repurposed for non-software tasks like UK tax analysis by using a folder of documents as a knowledge base.
- The process involves scraping tax legislation and manuals from gov.uk, resulting in ~10,000 documents for the agent to search.
- A CLAUDE.md file provides instructions for the agent, including creating parallel subagents for comprehensive research.
- Testing on ATT exam papers showed strong performance, with the agent scoring 2.5 out of 3 marks, outperforming regular LLMs.
- Potential applications include contract-invoice cross-referencing and content library management, limited mainly by the terminal interface.
- Future developments include GUI-based solutions and broader use in professional services due to the low barrier to creating expert systems.
How I use Claude Code to manage sysadmin tasks
2 months ago
- Use individual folders with CLAUDE.md files for task sets, versioned separately in Git, to manage sysadmin tasks with Claude Code.
- Structure folders by project and task group (e.g., app maintenance, DBA) including server info, project context, and common tasks for progressive disclosure.
- Enhance security by using SSH keys, bastion hosts, and ~/.ssh/config aliases to avoid storing credentials in repos.
- Pair rigorously with Claude Code: research each step, ask for plans and verification, and review every SSH command before execution.
- Document processes in CLAUDE.md after tasks (e.g., backups) to create self-documenting playbooks with verification steps and known issues.
- Extend the approach to cloud-first environments using CLI tools (e.g., gcloud, aws) with proper access management.
- Separate sysadmin CLAUDE.md from code repos to avoid context confusion and allow overarching infrastructure views.
- Update CLAUDE.md regularly to accumulate resolved issues and reflections, improving future resolution time and documentation.
- Make CLAUDE.md human-readable—summarize or convert to HTML for team sharing and cross-server queries.
No, it doesn't cost Anthropic $5k per Claude Code user
2 months ago
- Forbes claimed Claude Code Max users cost $5,000 in compute, but this is based on retail API prices, not actual inference costs.
- Actual compute costs are roughly 10% of API prices, as shown by open-weight models on OpenRouter (e.g., Qwen 3.5 at $0.39/MTok input vs Opus $5/MTok).
- Heavy users cost Anthropic ~$500 per month, not $5,000—a $300 loss, not $4,800.
- Most users are break-even or profitable; average Claude Code user consumes $180/month API-equivalent, costing $18 to serve.
- Cursor faces the real $5,000 cost because they pay retail API prices for Anthropic models.
- The narrative that AI inference is a money pit is misleading; it discourages competition and justifies high API markups.
I Sold Out for $20 a Month and All I Got Was This Perfectly Generated Terraform
2 months ago
- The author initially dismissed LLM tools like Copilot and Gemini as worthless and unethical, comparing them to overhyped trends like the Metaverse and NFTs.
- Using Claude Code changed his perspective: it effectively handled tedious programming tasks (e.g., Terraform, READMEs, YAML, GitHub Actions), saving time and reducing resentment.
- He acknowledges the ethical problems with LLMs—stealing human knowledge and benefiting the wealthy—but finds the tool too useful to ignore.
- Pro-LLM arguments include dismissing IP/copyright as irrelevant and prioritizing feature delivery over code quality, as exemplified by a pragmatic friend who succeeds with minimal craftsmanship.
- The author struggles with the moral compromise: using a tool that is both effective and unethical, feeling like a 'mercenary' rather than an 'artist' in his work.
Switching to Claude Code + VSCode inside Docker
2 months ago
- The author transitioned from using ChatGPT Plus and Cursor Pro to running Claude Code in Docker via VSCode's Dev Container feature.
- Cursor's recent rate limits and slow response times (especially with Sonnet 4) frustrated the author, leading to a switch to Claude's $20 subscription for better value and faster responses.
- Security concerns with AI agents (e.g., accidental file deletion, secret exposure) motivated running Claude Code in a containerized, isolated environment.
- The Docker setup restricts Claude to only access files inside the container, preventing damage to the host system and local secrets.
- A quick guide is provided: clone the repo, open in VSCode, click 'Reopen in Container', and Claude Code is ready to use.
- For git operations, using a fine-grained GitHub token with read/write access to 'Contents' is recommended over SSH keys to limit permissions.
- The token allows basic git commands (clone, pull, push) without granting excessive account access, enhancing security.
Claude Code uses Bun written in Rust now
2 months ago
- Jarred Sumner claimed Claude Code v2.1.181+ uses the Rust port of Bun, resulting in 10% faster startup on Linux.
- Evidence from strings command shows Bun v1.4.0 (a preview version) and 563 .rs source files embedded in Claude Code.
- The Rust version of Bun is now available as a canary release via `bun upgrade --canary`.
- A simple script can confirm the embedded Bun version by running `BUN_OPTIONS="--preload=/tmp/bun-version.ts" claude --version`.
- The commit updating Bun's version to 1.4.0 was made on May 17th, but hasn't been tagged in a stable release.
A Fireside Chat with Cat and Thariq from the Claude Code team
2 months ago
- Claude Tag lands 65% of product engineering PRs for the Claude Code team at Anthropic.
- Claude Code features are first tested internally and only shipped if they demonstrate user retention.
- System prompts for frontier models (e.g., Opus 4.8, Fable) have been reduced by 80%; giving examples and 'don't do' lists is no longer best practice.
- Auto mode in Claude Code is heavily evaluated and enables safe, long-running agent tasks; it also powers Claude Tag.
- Critical code changes are still manually reviewed, but automated code review handles outer layers after building trust through evals.
- Fable can one-shot many features including editing video, and performing complex multi-step tasks.
- To offset the feeling of loss from coding automation, Thariq advises being more ambitious with the work you take on.
- Anthropic's culture of working in public (e.g., public Slack channels) and 'ant fooding' (dogfooding) are key to successful tool adoption.
- Non-engineers use Claude Tag for searching company knowledge, checking metrics, and asking about features without coding.
Claude Code Deletes Your Context History from Your Device After 30 Days
2 months ago
- Consumer users (Free, Pro, Max) can choose to allow data for model training; commercial users (Team, Enterprise, API, etc.) have data not used for training unless explicitly opted in via programs like the Development Partner Program.
- Feedback via /feedback, /bug, or /share commands sends conversation transcripts to Anthropic (retained 5 years) for non-third-party users; for third-party providers, data is saved locally.
- Session quality surveys record only ratings; optional transcript sharing with redaction of keys/tokens, retained up to 6 months; can be disabled via environment variables.
- Data retention: consumer users who allow training have 5-year retention; otherwise 30 days; commercial users have 30 days or zero data retention for qualified Enterprise accounts.
- Telemetry includes usage metrics (no code/prompts/paths) and error reports (redacted secrets); both can be individually disabled via environment variables.
- WebFetch domain safety check sends only hostname to Anthropic for blocklist, unaffected by non-essential traffic settings; can be disabled in settings.
Beyond grep: The case for a context-rich AI coding harness
2 months ago
- Recent AI coding developments focus on software managing models, not just the models themselves.
- Anthropic's Claude Code team maintains a 'lean harness' that avoids over-planning or limiting features due to rapid model improvements.
- A harness is the software layer between AI models and developer projects, controlling model inputs, actions, and code base interactions.
- Claude Code trusts models to guide future development rather than imposing rigid opinions.
- Other harnesses include OpenAI's Codex, Google's Antigravity, open-source OpenCode, and startup options like Cursor or Augment Code.
- Claude Code deliberately avoids pre-building structured context around codebases by default, as evaluations show no measurable benefit.
Claude Code uses Bun written in Rust now
2 months ago
- Jarred Sumner announced Claude Code v2.1.181+ uses Rust port of Bun, speeding up Linux startup by 10% with little user notice.
- Evidence found in Claude Code installation shows Bun v1.4.0 (macOS arm64), a version not yet released on GitHub as of the text's date.
- Rust version of Bun is available as a canary release, and 563 Rust source files in Claude indicate it's running in production.
Setting up your spare Mac for Claude Code to control, a step-by-step guide
2 months ago
- Set up a spare Mac as a dedicated machine for Claude Code with computer use enabled to isolate risks.
- Configure SSH access and passwordless sudo on the target Mac for remote administration.
- Disable sleep and screensaver to keep the machine always-on.
- Install Claude Code on the target Mac with additional environment setup tools.
- Enable computer use functionality by installing a tmux server with GUI permissions.
- Use convenience scripts like ic.sh and clip.sh for remote session management and clipboard sharing.
- Set up optional VPN, Screen Sharing, and Tailscale for remote access and network security.
Claude Code: Anatomy of a Misfeature
2 months ago
- Anthropic introduced a 60-second auto-continue feature in Claude Code version 2.1.198 that allowed agents to proceed without user input after a timeout, presenting risks for unattended operations and deployments.
- The feature was shipped without any mention in the changelog or documentation, surprising users and raising concerns about safety, token waste, and the integrity of explicit user decision gates.
- After user reports and backlash, Anthropic quickly reversed the default behavior in version 2.1.200, making auto-continue opt-in via configuration, but the underlying code remained unchanged.
- Investigation revealed the feature was deliberately designed, instrumented with analytics, and included UI elements, yet was never officially announced or documented upon release.
- The incident highlighted issues with Claude Code's auto-update mechanism and the lack of public source code or transparent release processes, undermining user trust in the product's update governance.
Claude Code May–July 2026 weekly limits promotion
2 months ago
- Weekly usage limits in Claude Code are increased by 50% from May 13, 2026 through July 19, 2026.
- Promotion applies automatically to Pro, Max, Team plans, and legacy seat-based Enterprise users; Free plans and consumption-based Enterprise seats are excluded.
- Increase applies to Claude Code usage across CLI, IDE extensions, desktop, and web; 5-hour usage limits remain unchanged.
- No action is required to participate; users can check updated limits via the CLI /usage command.
- After July 19, 2026, weekly usage limits return to standard levels with no changes to plan or billing.
Show HN: Code-Manager
3 months ago
- Browse, preview, edit, create, rename and locate the entire Claude Code user directory with safe boundaries.
- Manage provider configurations, models, environment variables, permissions, sandbox, hooks, plugins, and statusline with preview and diffing capabilities.
- Scan token usage and cost from project files, aggregate by date, project, session, and model with SQLite caching and optional desktop widget.
- Manage user-level memory files like CLAUDE.md and rules/*.md, including built-in presets and import functionality.
- Create, edit, enable, disable, and delete Claude Code skills, with support for symlinked skills and one-click sync.
- View history details and heatmaps by project/session, and read local statistics and recent sessions.
- Show project Git status, branches, worktrees, project-specific files, skills pairing, and project auto-memory; includes terminal/editor opening and cleanup.
- Display active profile and live session status in the tray with count styles and pending pulse; supports macOS terminal refocus and LED device mirroring.
- Configure language, theme, default terminal/editor, launch-at-login, pricing toggles, focus shortcuts, and diagnostic tools like log viewer and system info.
Anthropic embedded spyware in Claude Code – and attempted to hide it from you
3 months ago
- Since version 2.1.91, Claude Code checks if a proxy is enabled and covertly transmits data via invisible system prompt alterations, including detecting Chinese timezones, proxy URLs, and affiliations with Chinese AI labs.
- Anthropic uses subtle changes in date formats and apostrophes in the system prompt to embed this information, potentially for detecting unauthorized resale and distillation in China, while obfuscating the code in the binary.
- This covert transmission is criticized as a violation of user trust, raising concerns about potential misuse, such as remote code execution or data exfiltration, given the extensive system access developers grant Claude Code.

firstprev0next

claude code

Show HN: Galda – Come back to Claude Code/Codex without digging through Git

Boris Cherny on Trying to Get Claude Code to Rewrite the Claude App

An Android and iOS app from 7 Claude Code commands, every prompt and timing

Claude Code system prompt says not to ask for permission, assumes user is absent

Claude Code

I gave Claude Code a folder of tax documents and used it as a professional tax agent

How I use Claude Code to manage sysadmin tasks

No, it doesn't cost Anthropic $5k per Claude Code user

I Sold Out for $20 a Month and All I Got Was This Perfectly Generated Terraform

Switching to Claude Code + VSCode inside Docker

Claude Code uses Bun written in Rust now

A Fireside Chat with Cat and Thariq from the Claude Code team

Claude Code Deletes Your Context History from Your Device After 30 Days

Beyond grep: The case for a context-rich AI coding harness

Claude Code uses Bun written in Rust now

Setting up your spare Mac for Claude Code to control, a step-by-step guide

Claude Code: Anatomy of a Misfeature

Claude Code May–July 2026 weekly limits promotion

Show HN: Code-Manager

Anthropic embedded spyware in Claude Code – and attempted to hide it from you