TLDR DEV 2025-12-22

Schedule a demo to learn more

PERFORMANCE HINTS

Jeff Dean and Sanjay Ghemawat, some of Google's most important engineers for over 25 years, wrote a guide on optimizing software performance. Performance needs to be considered early in critical code paths instead of solely relying on post-development profiling for optimization. The guide also covers algorithmic improvements, efficient memory representation, reducing allocations, avoiding unnecessary work, and using parallelization.

INTRODUCING CSS GRID LANES

CSS Grid Lanes is a new web standard to natively create flexible, dynamic masonry-style layouts on the web without JavaScript. This feature uses `display: grid-lanes` along with existing CSS Grid properties like `grid-template-columns` or `grid-template-rows` to define the layout's structure. It supports advanced capabilities such as variable lane sizes, item spanning, explicit placement, and introduces a "tolerance" property for fine-tuning item packing order.

CODE REVIEW AS HUMAN ALIGNMENT, IN THE ERA OF LLMS

Code review within teams should change to be primarily a human alignment mechanism. Large language models have no memory and have to rediscover everything in the system from first principles every time they are run. Humans can collectively maintain a shared vision of what the system should do. Code review should reconfigure itself around this purpose.

MORE DATABASES SHOULD BE SINGLE-THREADED

From a DX perspective, every transaction within a shard would always be serializable and never needs to be retried, making it much easier to think about. A single-threaded database would be fully horizontally scalable and have extremely high per-thread throughput, as it wouldn't need to sync with other threads. It would be a much more predictable system. While more restrictive, there would be no more hidden race conditions.

Planview Hub Bytes is a series of 15-minute expert-led sessions

covering the most requested integrations. Watch solution architects demo Jira + ServiceNow cross-functional workflows, no-code integration setup, and connecting custom in-house tools. After the quick demo, you'll have a live Q&A with architects who've deployed these at

hundreds of organizations. Register for the series

CLAUDE'S FRONTEND DESIGN SKILL

This gist contains a Claude skill for creating production-grade frontend interfaces that avoid generic AI slop aesthetics. The user provides frontend requirements, and the AI generates production-grade and functional code with visually striking and memorable design. The skill instructs the agent to never use generic AI-generated aesthetics, to interpret prompts creatively, and to make unexpected choices that feel genuinely designed for the context.

DESTRUCTIVE GIT COMMAND PROTECTION FOR CLAUDE CODE

AI agents can execute destructive commands without understanding the consequences. They can do this even if the AGENTS.md file forbids such commands. This post provides a Claude Code hook that blocks certain git commands before they can run. The hook uses regex pattern matching, so clever or obfuscated commands may be able to bypass it.

GARAGE

Garage is a lightweight, efficient, and resilient S3 object store designed for reliability. Capable of running outside traditional datacenters, it guarantees data redundancy through 3-zone replication. Garage is a self-contained, dependency-free binary that is fast to deploy on various Linux distributions and machines.

TRON

Lite³ (TRON) is a zero-copy, schemaless binary serialization format that organizes data as a B-tree within a single contiguous buffer. This design allows for direct, O(log n) access and mutation of data fields without traditional parsing or serialization steps. It supports conversion to/from JSON, but is much more performant.

RUE

Rue is a memory-safe language that is higher-level than Rust, but lower-level than Go. It features a simple syntax inspired by various programming languages. Rue compiles directly to native code. It is still in early development.

2025 LLM YEAR IN REVIEW

2025 was an eventful year for LLMs, with paradigm shifts that redefined the field. Reinforcement Learning from Verifiable Rewards (RLVR) came up as a new training stage, allowing LLMs to develop reasoning strategies and driving much of the year's capability progress. The year also saw the rise of new application layers, local AI agents like Claude Code, "vibe coding" that democratized programming, and early hints of a visual LLM GUI.

BACKING UP SPOTIFY

Anna's Archive has created the world's first preservation archive for music by backing up Spotify's metadata and music files. This massive 300TB collection covers 256 million tracks and 186 million unique ISRCs, representing approximately 99.6% of all Spotify listens. The project archives a "long tail" of less popular music and makes the entire archive openly mirrorable.

CODE ORANGE: FAIL SMALL — OUR RESILIENCE PLAN FOLLOWING RECENT INCIDENTS

A 'Code Orange' at Cloudflare is an incident that requires top priority from everyone across the company, with teams working cross-functionally as necessary to get the job done while pausing any other work.

JSON STRUCTURE: A JSON SCHEMA LANGUAGE YOU'LL LOVE

JSON Structure is a strictly typed data definition language that describes JSON-encoded data so that mapping to and from programming languages and databases becomes straightforward.

BIG GPUS DON'T NEED BIG PCS

A Raspberry Pi 5 can use powerful GPUs for demanding tasks like AI and media transcoding, often achieving performance within a few percent of a desktop PC.

WHY I DON'T LIKE “CORRECTION OF ERROR”

Amazon's term for its incident review process, Correction of Error, implies that incidents occur as a result of errors, which limits the benefits available from the incident review process.

WORDIEST

This developer used GPT-5.2 to port an old Android game to iOS - the resulting app even got approved on the App Store.