Coding agents · 23 min readWhat is spec-driven development and when does it help?The agent writes a spec, a plan and a task list before the code. We cover the tools, the criticism and reviewing against a spec.Read the post AI security · 22 min readWhat is MCP security and what to check in a pull requestAn MCP server in a repository runs on every approver’s machine. We read five tools’ docs and wrote a pull request checklist.Read the post Coding agents · 18 min readWhat is git worktree and how do parallel agents use it?A git worktree gives each coding agent its own files, but not everything is separate. We checked five tools and our repository.Read the post Coding agents · 20 min readWhat are agent skills, and how do you write SKILL.md?A skill is a folder the agent reads only when it needs it. Where to put one in Claude Code, Codex, Cursor, Gemini CLI and Copilot.Read the post Building AI agents · 10 min readWhat is LLM-as-a-judge, and when can you trust the judge?LLM-as-a-judge means having one language model grade another model’s answers. Here is how we audited our own judge model.Read the post Code review · 12 min readHow to review AI-generated codeAI-generated code compiles, passes its tests and can still be wrong. Five questions to ask of a pull request, and a 15-point checklist.Read the post Coding agents · 19 min readWhat are subagents and how do you use them in five coding tools?In Claude Code, Codex, Cursor, Gemini CLI and Copilot, a subagent works in its own context and reports back. Setup, limits and best practices, from the docs.Read the post Coding agents · 15 min readWhat are hooks in coding agents, and how do you write them?A coding agent doesn’t always follow a rule in its rules file, but a hook runs every time. Here is how to write hooks in five tools, with examples.Read the post Building AI agents · 18 min readPrompt caching: 8 best practices for Claude, OpenAI and GeminiPrompt caching serves a prompt’s repeated prefix from a cache. Here is how to turn it on at four providers, how long it stays warm and how to raise the hit rate.Read the post Code review · 11 min readHow to write a code review prompt: six rules and a templateSix rules for writing a code review prompt for Claude Code, Codex, Copilot and Cursor, plus a template you can copy and adapt.Read the post Building AI agents · 17 min readHarness vs loop engineering vs context engineeringHarness engineering shapes what surrounds an AI agent’s model; loop engineering builds the system that keeps running it. Definitions, sources, an example.Read the post Coding agents · 14 min readHow to write AGENTS.md and CLAUDE.md: what 100 repos doAGENTS.md and CLAUDE.md give coding agents a project’s rules. We read the 100 most-starred repos: contents, length, linking.Read the post Code review · 12 min readBlast radius: what a pull request touches outside its diffBlast radius is the code a change reaches beyond its diff. Using a real pull request, we show how to find it from the code alone, without drawing false arrows.Read the post Code review · 14 min readStacked pull requests: how to set up and review a stack on GitHubStacked pull requests split a large change into small layers. We look at GitHub’s built-in support and at how to review code in a stack.Read the post diffox · 11 min readHow we use Jev for two jobs in AI code reviewJev, a decision model, answers with numbers, not text. We use it for two jobs, and our measurements give one rule: it compares, it does not judge.Read the post diffox · 12 min readdiffox architecture: 17 specialist agents, four stages, one reviewdiffox splits an AI code review across 17 specialist agents. Every finding comes with a quote from the code, and code, not a model, checks that quote.Read the post diffox · 18 min readdiffox is the cheapest of the top five by score: $0.41 per reviewWe compared 23 AI code review tools on Martian’s independent benchmark. diffox is the cheapest of the top five by score: $0.41 a review; the other four $1.50–5.63.Read the post diffox · 10 min readdiffox ultra ranks first among 23 AI code review toolsWe ran Martian’s open-source code review benchmark ourselves. diffox ultra caught 119 of 158 real bugs and ranked first. Every review is public.Read the post
