Developer Tools: Automation

Browser and file control are becoming precise, code-driven primitives, and interactive HTML has joined as a new output format: Claude Code's interactive HTML skill launches a localhost server that allows in-browser commenting on generated pages, with automatic content updates and guided tours of changes — a Google Docs-style iterative workflow for HTML outputs.: WebMCP controls the main Chrome instance without sandboxed Playwright, dev-browser lets agents drive browsers by writing code (faster than screenshot-based computer use), agent-browser uses 82% fewer tokens than Playwright MCP, and MarkItDown converts any file (PDF/PPTX/DOCX/Excel/audio/YouTube) to clean Markdown because LLMs reason better on the format they were trained on. Server-side automation is unlocked (Obsidian Headless Publish/Sync without the desktop app), local-first ingestion runs without data leaving the machine (brain-ingest extracts 12-18 claims from a 90-minute talk), and Obsidian gets fuzzy + precise retrieval via smart-connections + qmd MCP servers. Issue tracking is going real-time and agent-driven: LogRockets surfaces customer issues for a markdown-based fix workflow without planning decks, and Symphony assigns a Codex agent to every open issue — turning trackers into always-on systems where humans shift from doing to reviewing. Productivity skills give agents eyes, self-reflection, and self-improvement (/ss screenshot skill saving ~1 hour/week, Codex's hidden Chronicle analyzing usage patterns, claude-smart turning repeated corrections into reusable local rules), while structured /goal prompts add measurable outcomes, constraints, verification, rollback, and stop rules to agent tasks. The pattern replicates across verticals — ml-intern automates the post-training research loop end-to-end (beating Claude Code on GPQA), finance has its own dedicated Ai Trading sub-ecosystem, and systematic diagnostics now treat before/after tests as the closure condition for mundane ops work. The supporting cast (Goose, Paperclip Desktop, Orca IDE's Notion-style editing, official Claude Code setup plugins) and a hard scaling lesson (git's 5GB ceiling forces SQLite migration for 2.3GB+ knowledge bases) round out the agent-automation tooling layer.