How to add real images and icons to a project by pulling them from a source — stock photos via a provider API (Pexels default, Unsplash swap) behind one swappable constant, and icons via Iconify/astro-icon (Lucide) — with a dormant-safe fetch-script pattern, Astro src/assets optimization, and an images[] content field. Use when a site/blog needs real photos or proper SVG icons instead of emoji/gradient placeholders, or when wiring per-post imagery.
Defensive patterns for Chrome MV3 content scripts that have to survive host-app surprises — extension reloads orphaning the chrome.* namespace, host frameworks recycling DOM elements between component instances. Use when writing or debugging any content script that injects into a third-party page or holds long-lived observers/listeners.
Rules for organizing assets in a Chrome MV3 extension repo and the Chrome Web Store submission asset checklist. Use when adding icons or promo art, packaging a build for upload, or auditing whether a repo is ready for store submission.
Gotchas for driving the `claude` CLI headlessly (`-p`/`--print`) from shell scripts, hooks, and crons, AND for writing Claude Code lifecycle hooks safely. Pass the prompt via stdin — the variadic `--allowedTools <tools...>` silently eats a trailing positional prompt and aborts the CLI. SessionStart fires repeatedly (startup/resume/clear/compact) so guard GUI/side-effects to once-per-boot or they can loop. Use when writing or debugging any script, SessionStart/SessionEnd/PreCompact/UserPromptSubmit hook, or cron wrapper (nervepack's own episodic capture/maintain/evaluator/recall hooks and the dashboard-launch hook all rely on this).
Framework-level UI implementation gotchas (Preact/React and similar) that bite repeatedly — focusing/selecting dynamically-mounted inputs, making inline-edit commit exactly once, and silencing ResizeObserver loop warnings. Use when building editable/inline-rename fields, doing focus management, wiring a ResizeObserver, or debugging "the input won't focus", "typing doesn't replace the text", "it saved twice / didn't save", or "ResizeObserver loop completed with undelivered notifications", or a CSS bar/progress "fill" that renders invisibly because an inline element ignores width/height.
Gotchas when committing/inspecting changes with git via a shell (esp. the Bash tool) that lead to wrong commits or wrong conclusions. Chiefly — a file MOVE split across two commits is NOT detected as a rename, so a `D old/path` in `git status` can be a move (content still exists elsewhere), not a deletion; and backticks/`$(...)` inside a double-quoted `git commit -m "..."` trigger shell command substitution and silently mangle the message. Use before writing a commit message that says "deleted/removed/retired", before `git add -A` in a dirty/reorg tree, or when a commit message came out garbled.
Gotchas for authoring portable Mermaid diagrams that render correctly on GitHub and in editor/VSCode previews — not just in an online validator. Use when writing or embedding a Mermaid diagram in Markdown, or debugging one that renders as raw text, shows literal `&entity;` codes in labels, or fails to draw on GitHub/VSCode even though mermaid.live says it's valid.
Gotchas when parsing/ingesting Markdown that was EXPORTED from a rich editor (Google Docs, Confluence, Notion, Word→MD) rather than hand-written. Chiefly: the exporter backslash-escapes Markdown special chars, so a regex that matches before unescaping silently truncates names and collapses distinct entities. Use when writing a parser or ingestion adapter over human-authored `.md` docs (data-model docs, wiki exports, spec dumps) — especially if some tables/sections come out empty or duplicated.